Decidable and Undecidable Problems in Matrix Theory

Vesa Halava
University of Turku, Department of Mathematics, FIN-20014 Turku, Finland vehalava@utu. Supported by the Academy of Finland under the grant 14047

Turku Centre for Computer Science TUCS Technical Report No 127 September 1997 ISBN 952-12-0059-6 ISSN 1239-1891

Abstract
This work is a survey on decidable and undecidable problems in matrix theory. The problems studied are simply formulated, however most of them are undecidable. The method to prove undecidabilities is the one found by Paterson Pat] in 1970 to prove that the mortality of nitely generated matrix monoids is undecidable. This method is based on the undecidability of the Post Correspondence Problem. We shall present a new proof to this mortality problem, which still uses the method of Paterson, but is a bit simpler.

Keywords: decidability, undecidability, matrix semigroups, mortality, freeness, niteness, zero in the right upper-corner, Skolem's problem

Contents
2.1 2.2 2.3 2.4 2.5 2.6 2.7

1 Introduction 2 Preliminaries

Basics . . . . . . . . . . . . . . . . Semigroups and monoids . . . . . . Matrix semigroups and monoids . . Decidability and undecidability . . Word monoids and morphisms . . . The Post Correspondence Problem The Mixed Modi cation of PCP . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

1 3
3 3 3 6 7 8 8

3 Mortality of Matrix Monoids

3.1 3.2 3.3 3.4

The undecidability of the mortality problem . . . . . . The mortality of semigroups with two generators . . . The existence of the zero element in matrix semigroups The mortality problem in dimension two . . . . . . . .

11
11 15 17 18

4 Freeness of Matrix Semigroups

4.1 The undecidability of the freeness . . . . . . . . . . . . . . . . 21 4.2 The freeness problem in dimension two . . . . . . . . . . . . . 23 4.3 Common elements in semigroups . . . . . . . . . . . . . . . . 24 5.1 The decidability of the niteness . . . . . . . . . . . . . . . . . 25

21 25

5 Finiteness of Matrix Semigroups 6 Zero in the Right Upper Corner 7 Skolem's Problem

6.1 The undecidability of the zero in the right upper corner . . . . 33 6.2 The two generator case . . . . . . . . . . . . . . . . . . . . . . 35

33 37

7.1 Skolem's problem in dimension two . . . . . . . . . . . . . . . 37 7.2 Skolem's problem and the mortality problem . . . . . . . . . . 42 7.3 Skolem's problem for matrices over natural numbers . . . . . . 45

8 An In nite Number of Zeros in Skolem's Problem 9 Summary

8.1 Rational series . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 8.2 Skolem's theorem . . . . . . . . . . . . . . . . . . . . . . . . . 51 8.3 Decidability of the in nite number of zeros . . . . . . . . . . . 58

46 60

1 Introduction
This thesis deals with decidable and undecidable problems in matrix theory. The matrices we consider are over integers, and the problems we study are simply formulated and elementary, but it turns out that many of these problems are algorithmically undecidable, which means that there exists no computer program to solve these problems. We use mainly only basic properties of matrices. These and other basic de nitions, notations and properties are presented in the next chapter. Occasionally we need more advanced properties of matrices that are stated when they are needed. The de nitions and proofs of such properties can be found from Gan]. We are going to show that many simply formulated problems are actually undecidable. Our undecidability proofs are based on the fact that the so called Post Correspondence Problem is undecidable. This undecidable problem was introduced by E. Post Pos] in 1946 and it is a very important problem in formal language theory. As we shall see, it is also suitable for our purposes in proving undecidability results. In the third chapter we study the existence of the zero matrix in matrix semigroups, i.e. the so called mortality problem. The proof of the undecidability of the mortality problem for 3 3 matrices by M.S. Paterson Pat] in 1970 provides a tool to prove other undecidability results for matrices. Paterson used the Post Correspondence Problem in his proof, and throughout this work we shall use the very same method. We shall present a shorter and simpli ed proof for the undecidability of the mortality problem, however, using the same method than the original proof of Paterson. We shall also consider the special case where the semigroup is generated by two matrices. It turns out that also in this case the mortality problem is undecidable if the dimension of the matrices is at least 45. This result is from CKa]. In the third chapter we shall show that the method of Paterson does not suit for problems of 2 2 matrices. This result is from CHK]. It follows that for many problems, when we consider such matrices, it is not known whether they are decidable or not. 1

In the fourth chapter we study the freeness of matrix semigroups. This problem is important from the point of view of semigroup theory, especially for 2 2 matrices. Results in that chapter is mainly from CHK]. In the fth chapter we consider the niteness of matrix monoids. We will prove that it is decidable whether a matrix semigroup of matrices over natural numbers is nite or not. The proof we present is from MaS]. The decidability of niteness helps us also to prove some other simple properties of matrix semigroups generated by one matrix. These results are mentioned at the end of the chapter. In the sixth chapter the existence of a zero in the right upper corner in a matrix semigroup is treated. We shall prove that this problem is undecidable for matrices with dimension at least 3. The proof we present is from Man]. We shall also consider the case where semigroup is generated by two matrices and we prove that this problem is undecidable, when the dimension of the matrices is at least 24. This result is from CKa]. In the seventh and eighth chapters we study the existence of a zero in the right upper corner in a matrix semigroup generated by one matrix. This problem is so called Skolem's problem. In the seventh chapter we will show that Skolem's problem is decidable for 2 2 matrices. The proof we present is based on ideas of J. Cassaigne Cas]. In the eighth chapter we will prove that it is decidable, whether there is in nite number of zeros in the right upper corner in the powers of one matrix. This proof, due to Han], is elementary, but not simple and it uses the theory of rational series. The proof is based on so called Skolem's theorem. Results of rational series in that chapter is from BeR]. We shall summarize the results of this work in the tenth chapter where a table of decidable and undecidable matrix problems is presented.

2 Preliminaries
2.1 Basics
We shall denote the set of natural numbers by N . It is assumed troughout this paper that zero is in N , i.e. N = f0; 1; 2; : : : g. The sets of integers, rational numbers, real numbers and complex numbers are denoted by Z, Q , R and C , respectively.

2.2 Semigroups and monoids

Let S be a set. A pair (S; ) is called a semigroup, if is a binary operation such that, for all a, b and c in S , a b 2 S and (a b) c = a (b c): In other words is an associative operation in S . Usually the operation is called a product and the semigroup (S; ) is denoted simply by S . Semigroup S is called a monoid, if it has a neutral or identity element, i.e. there exists an element 1 in S such that for all a in S , 1 a = a 1 = a: Form now on, we denote ab = a b if no misunderstanding is possible. A semigroup S is said to be freely generated and it is free if there exists subset X of S such that every element of S has a unique factorization over X , i.e. every element of S can be uniquely expressed as a product of elements of X . In this case the set X is called the free generating set. A monoid M is said to be free, if M nf1g is a free semigroup.

2.3 Matrix semigroups and monoids

In this work we consider square matrices over integers, and in this section we shall recall some basic notations and properties of these matrices. As 3

a general reference for matrix theory we give F.P. Gantmacher's book The Theory of Matrices Gan]. The set of n n matrices over integers is denoted by Zn n, and n is called the dimension. Sometimes Z is replaced by N or by some other semiring. Recall that K is a semiring, if it has two operations, + and , satisfying the following properties: (i) (K; +) is a commutative monoid with 0 as its neutral element. (ii) (K nf0g; ) is a monoid with 1 as its neutral element. (iii) For all a; b; c in K , a(b + c) = ab + ac and (b + c)a = ba + ca. (iv) For all a in K , 0a = a0 = 0. Let A be a matrix in K m n . We denote the coe cient in the i'th row and on the j 'th column of matrix A by Aij , and, consequently, a matrix A is de ned to be (Aij )m n . It is clear that (Zn n; ) forms a monoid, where is the usual multiplication of matrices, i.e. if A and B are in Zn n, then (A B )ij = (AB )ij =
n X k=1

Aik Bkj ;

and the identity element is the identity matrix I (or In), 01 0 0 : : : 01 C B I = B0: : 1: : 0: : :: :: :: : :0C : @: : : :A 0 0 0 ::: 1 n n We will denote this monoid by Mn(Z). The above can be extended to other semirings than Z, and the matrix monoid of n n matrices over semiring K is denoted by Mn(K ). We call a matrix P in Mn(N ) a permutation matrix, if it is derived from In by mixing the rows (or columns) by some permutation. Clearly, if P is a permutation matrix, then each row and each column has 1 in exactly one position and in the other positions there are 0's. We denote ( 1; if i = j; ij = 0; otherwise: Clearly Iij = ij , and if is a permutation on the set f1; : : : ; ng, then P , for which Pij = i; (j), is a permutation matrix. It is also easy to see that P ?1 = (i);j . 4

For a square matrix A = (Aij )n n de ne the determinant of A by

A11 A12 : : : A1n det(A) = :A21: : :A22: : ::::::: : :A2:n: :: :: : An X1 An2 : : : Ann = sign(j1 ; j2 ; : : : ; jn)A1j1 A2j2

(2.1)

Anjn ;

where runs through the permutations on f1; 2; : : :; ng, = (j1 ; j2; : : : ; jn), and sign(j1 ; : : : ; jn) is the sign of the permutation , sign(j1 ; j2; : : : ; jn) = (?1)t( ) ; where t( ) is the number of inversions in , i.e the number of cases where jl > jk and l < k. Recall that in a permutation (j1 ; : : : ; jn) all ji's are di erent. There are also some other methods than the sum (2.1) to calculate the determinant, but we will use the one in the de nition. Let A be an element of Mn (Q ). If det A 6= 0, we say that A is invertible, and it has the unique inverse matrix A?1 in Mn (Q ) such that AA?1 = A?1A = I . Note that if A is an element of Zn n and det(A) 6= 0, then the inverse matrix of A is not necessarily an element of Zn n, but it is an element of Q n n . The eigenvalue in C and eigenvector x in C n , x 6= (0; : : : ; 0), of a matrix A in Mn (C ) satisfy the condition

Ax = x;

(2.2)

where C n denotes as usual the n dimensional vector space over the complex numbers. The vector x can be understood also as an n 1 matrix. From the equation (2.2) it follows that the eigenvalue of A satis es the equation det(A ? I ) = 0: This equation is called the characteristic equation of the matrix A. The left hand side of this equation is so called characteristic polynomial, which is denoted by

cA( ) = (?1)n( n ? c1

n?1 + ?

+ (?1)ncn):

This polynomial has degree n, if A is in C n n , and the coe cients ci are in C . If 0 is an eigenvalue of A, we know that cA ( 0 ) = 0. Therefore, if we let the eigenvalues of A be i, i = 1; : : : ; n, it follows that

cA( ) = (?1)n( ? 1)( ? 2) ( ? n); where the i's need not be di erent. Now we see that cn =
n Y i=1 i = c(0) = det(A) and c1 = n X i=1 i:

We call the sum of i's the trace of the matrix A, and denote it by tr(A). Let p be a polynomial with coe cients in C , p(x) = c1xk +c2xk?1 + +c1. Then for a matrix A in Mn (Z) we de ne

p(A) = c1Ak + c2 Ak?1 + + c1 I: Next we present the so called Cayley{Hamilton theorem. We do not present a proof here, a detailed proof can be found for example from Gan]. Theorem 2.1. Let A be a square matrix and let cA be its characteristic
polynomial. Then

cA(A) = 0;
where 0 = (0)n n is the zero matrix. Note that since we are able to compute the determinant of a square matrix, we are also able to compute the characteristic polynomial of a square matrix. We end this section with a notation for the transpose matrix AT of A, T) = A . (A ij ji

2.4 Decidability and undecidability

Let P be a well-de ned decision problem, i.e. a set of instances each of which either has or does not have a certain property. We say that the problem P is decidable, if there exists an algorithm, which for every correct input or an instance of P terminates and gives an answer \yes", if the input has the required property and answers \no" otherwise. If no such algorithm exists, then P is called undecidable. To prove that the problem P is decidable, we 6

must nd an algorithm which decides P , or show that such an algorithm exists. The usual method to prove the undecidability of a problem P is to reduce some already known undecidable problem Q to P . This means that, we transform e ectively an arbitrary instance q of Q to some instance p of P , so that the condition \p has the property of P " is equivalent with the condition \q has the property of Q". It follows that if the problem P is decidable, then the instances q of the problem Q can be solved and therefore the problem Q is decidable, which leads to a contradiction. We shall use a classical undecidable problem, the Post Correspondence Problem or one of its modi cations, in such reductions. These undecidable problems are introduced later in this chapter.

2.5 Word monoids and morphisms

Let be a nite set of symbols. We call an alphabet and its elements are called letters. A nite sequence of letters is called a word. Denote by + the set of all words over . The set + is a semigroup, the word semigroup, when the binary operation is de ned as the concatenation of words, which means that, if w1 = a1 : : : an and w2 = b1 : : : bm are in +, then w1 w2 = a1 : : : anb1 : : : bm : The result of the catenation is clearly in + and it is an associative operation. Let be the empty word and = + f g. Clearly is a free monoid generated by and is its identity element. A subset of a word monoid is called a language. Next we de ne a few properties of words. Let w = a1 : : : an be a word in so ai is in for all 1 i n. The length of the word w is n, and we denote it by jwj = n. A word u in is called a pre x of the word w, if there exists a word v in such that w = uv. Then the word v is called a su x of w.
+,

A mapping h from a monoid M1 to a monoid M2 is called a morphism if, for all u and v in M1 , h satis es the condition h(uv) = h(u)h(v): Note that in the right hand side of this equation the operation is the operation of the monoid M2 . 7

2.6 The Post Correspondence Problem

We shall next study the Post Correspondence Problem, PCP for short, introduced by E. Post Pos] in 1946. Let h and g be two morphisms from into . The equality set of h and g is the set E (h; g) = fw 2 + j h(w) = g(w)g: The Post Correspondence Problem asks to decide for a given pair (h; g) whether or not E (h; g) = ;. Elements in E (h; g) are called solutions of the instance (h; g) of PCP. This version of PCP is not the original version of the problem, but equivalent and more useful to our purposes. The size of an instance (h; g) of PCP is de ned to be the cardinality of the alphabet . We denote by PCP(n) the subproblem of PCP for instances of size at most n. The undecidability of PCP is a classical result in formal language theory, and it was rst proved by Post in Pos] in the general case, i.e. there does not exist any algorithm for solving all instances of PCP. It is also known that if n 2, then PCP(n) is decidable, and if n 7 then it is undecidable. The proof of the undecidability of PCP(7) can be found from MSe] and the proof of the decidability of PCP(n), for n 2, can be found, for example, from HaK]. For n greater than 2 and smaller than 7 the decidability status is open.

2.7 The Mixed Modi cation of PCP

There exists many modi cations of the Post Correspondence Problem, which have been proved to be undecidable. We shall next introduce one of these that is later used in the Chapter 4. The mixed modi cation of PCP, MMPCP for short, asks to determine for two given morphisms h; g : ! whether there exists a word w = a1 : : : ak with ai in and k 1, such that h1(a1 )h2 (a2) : : : hk (ak ) = g1(a1)g2(a2 ) : : : gk (ak ); (2.3) where, for each i, hi and gi are in fh; gg and, for some j , hj 6= gj . The word w satisfying the equation (2.3) is called a solution of the instance (h; g) of MMPCP. 8

We show that MMPCP is undecidable. The proof is from CHK], cf. also HaK]. Theorem 2.2. MMPCP is undecidable. Proof. We use the method, mentioned earlier in this chapter, to prove the undecidability. This requires to reduce PCP to MMPCP, that is, to transform an instance of PCP to that of MMPCP such that both of these have a solution simultaneously. Let (h; g) be an instance of PCP and assume that h; g : ! and that c, d and e are new letters not in . Further, let mappings l; r : ! ( fdg) be morphisms de ned as l(a) = da and r(a) = ad for all a in . Finally, for each a in we de ne two morphisms ha ; ga : ( fd; eg) ! ( fc; d; eg) by setting

ha(x) = l(h(x)); ha (d) = cl(h(a)); ha (e) = de;

ga(x) = r(g(x)) ga(d) = cdr(g(a)); ga (e) = e:

for all x 2 ;

Next we show that the instance (h; g) of PCP has a solution aw, w in , if and only if the instance (ha ; ga) of PCP has a solution dwe. This follows from the equations

ha(dwe) = cl(h(a))l(h(w))de = cl(h(aw))de

and

ga(dwe) = cdr(g(a))r(g(w))e = cdr(g(aw))e:

Now, since the pair (ha ; ga) can only have solutions of the form dwe, we conclude that it is undecidable whether for given morphisms h; g : ! the instance (ha ; ga) of PCP has a solution. Finally we show that the pair (ha ; ga), as an instance of PCP, has a solution if and only if the same pair has a solution as an instance of MMPCP. To simplify notation we denote h = ha and g = ga. If (h; g) has a solution as an instance of PCP, then it has a solution also as an instance of MMPCP, therefore the implication in one direction is clear. So assume that the pair (h; g) has a solution as an instance of MMPCP and let w = a1 : : : ak be a solution of minimal length. We claim that then also h(w) = g(w), i.e. w is a solution of instance (h; g) of PCP. In notations of (2.3), the minimality of w implies that h1 6= g1 and hk 6= gk , so by the de nitions of h and g, a1 = d and ak = e. We see also that ai 9

is not d or e if i = 2; : : : ; k ? 1, because otherwise there would be a shorter solution than w. We may assume, by symmetry, that h1 = h and g1 = g and we will show that hi = h and gi = g for all i = 1; : : : ; k. Assume the contrary. Then there must be the smallest t such that gt = h or ht = g. Consider the rst alternative. Then we have

g(a1 : : : at?1 )h(at ) 2 cd( d)+(d )+;

and so there is a pre x in the right hand side of (2.3) that ends with dd. But no pre x of the left hand side of (2.3) matches with this pre x, because if hi 6= g for all i, then h1(a1 ) : : : hk (ak ) 2 c(d )+, and if hi = g for some i, then there is a pre x which is in c(d )+ . Therefore we do not get two d's without having two consecutive letters of rst. This is a contradiction. In the second alternative, similarly, we get that

h(a1 : : : at?1 )g(at) 2 c(d )+( d)+;

i.e. the left hand side of (2.3) has a pre x that ends with two consecutive letters of . But no pre x of the right hand side of (2.3) matches with this pre x, because g(a1) : : : g(ak ) 2 cd( d)+, and if, for some i, gi = h, then there is a pre x which is in cd( d)+(d )+. Therefore we cannot have two consecutive letters of without having a factor dd rst. This contradiction proves the claim. We shall use MMPCP in the proof of the undecidability of the freeness of matrix monoids, for which it suits very well.

10

3 Mortality of Matrix Monoids

Let S be a given nitely generated submonoid of n n matrices from Zn n. `Given' here means that we are given a nite generator set of S . In this section we consider the mortality problem, which is de ned next.

Problem 1. Determine whether the zero matrix belongs to S . In other

words, determine whether there exists a sequence of matrices M1 ; M2; : : : ; Mk in S such that M1 M2 Mk = 0.

3.1 The undecidability of the mortality problem

We prove that the mortality problem is undecidable for 3 3 matrices. This result was rst proved by M.S. Paterson in 1970 Pat]. He used clever coding techniques to make a reduction to PCP. We use the same method, but our proof itself is simpler, althought the idea remains the same. The basic idea of the proof is to reduce an instance of PCP to the mortality problem. Let ? be an alphabet. We use an injective morphism from ? ? into N 3 3 to represent a pair of words in the multiplicative monoid of matrices. Then, of course, the product of matrices representing pairs (u; v) and (u0; v0) represents pair the (uu0; vv0). First we de ne notions needed in the proof. Let ? = fa1 ; a2; :::; ang and de ne a function : ? ! N by (ai1 ai2 :::aik ) =
k X j =1

ij nk?j and

( ) = 0:

We see that, for each word w, the function gives the value which w represents as an n-adic number, and because each natural number has a unique n-adic representation, must be injective. We also note that, for all u and v in ? , (uv) = njvj (u) + (v): 11 (3.1)

Next de ne a mapping : ? ! N 2 2 by (ai) = n 0 ; i 1 for all i = 1; 2; :::; n. Now if i; j 2 f1; : : : ; ng, we get that (ai) (aj ) = n 0 i 1 2 = nin+ j

n 0 j 1 0 = njai aj j 0 : 1 (ai aj ) 1

Clearly this can be extended for all w 2 ? , and therefore

jwj 0 (w) = n(w) 1 :

This can be proved by induction. We note that is a morphism from ? into N 2 2 , since if u and v are in ? , then
juj 0 jvj 0 (u) (v) = n(u) 1 = n(v) 1 njujnjvj 0 (u)njvj + (v) 1 =

njuvj 0 = (uv): (uv) 1

is also injective, because is injective and (w) is the (2; 1)-entry of the matrix (w). Next we de ne a monoid morphism : ? ? ! N 3 3 , where two copies of is applied simultaneously in 3 3 matrices,

0 njuj 0 01 (u; v) = @ 0 njvj 0A :

(u) (v) 1

(3.2)

The fact that is a morphism follows from the fact that is a morphism. The morphism is also doubly injective, which means that, if (u1; v1 )31 = (u2; v2)31 , then u1 = u2, and if (u1; v1)32 = (u2; v2)32 , then v1 = v2. Notice also that for the empty word we have ( ; ) = I3. We are now ready to prove the undecidability result.

Theorem 3.1. The mortality problem is undecidable for 3 3-matrices with

integer entries.

12

Proof. First de ne matrix A,

01 A = @?1 1

0 1 0 ?1A : 0 0 0

Let Y be the set of matrices 0p W (p; q; r; s) = @0 q and p; q; s and r are integers. then

0 0 r 0A ; where p; r > 0; q; s 0; s 1 If B and C are in Y , say

B = W (p1; q1 ; r1; s1) and C = W (p2; q2 ; r2; s2);

0 0 r1 r2 0A = W (p1p2 ; q1p2 + q2 ; r1r2 ; s1r2 + s2 ); q1p2 + q2 s1r2 + s2 1 then clearly BC is in Y . Since also I is in Y , Y is a monoid. Let L be a nitely generated submonoid of Y and S be the matrix monoid generated by fAg L. We notice that A2 = A, i.e. A is an idempotent, and that for all W (p; q; r; s) in L,

0 pp 1 2 BC = @ 0

AW (p; q; r; s)A = (p + q ? s)A:

Next we show that 0 2 S () 9W 2 L : AWA = 0:

(3.3) (3.4)

The reverse implication is trivial, because if AWA = 0 for some W 2 L, then 0 2 S. Assume now that 0 2 S . Since 0 2 L and for all k 1, Ak = A, there = must be product

AW1 AW2A AWt A = 0 (3.5) for some t 1 and Wj 2 L, for all j = 1; : : : ; t. If we assume that t > 1 and that t is minimal, then, because A is an idempotent, we have AW1 A AW2A AWt A = 0:
13

Now by equation (3.3) we get that, for some integer m, mAW2 A AWt A = 0: Now we have two cases; if m = 0, then AW1 A = 0, and if m 6= 0, then t is not a minimal. Both of these cases lead to a contradiction with the minimality of t and therefore there exists W in L, such that AWA = 0. Next we reduce an instance of PCP to the mortality problem using mapping from ? ? into N 3 3 as de ned in (3.2). Assume that ? = fa1; a2 ; a3g, so that n = 3. Let (h; g) be an instance of PCP, where h; g : ! and = fa2; a3 g. De ne the matrices Wa = (h(a); g(a)) and Wa0 = (h(a); a1g(a)) (3.6) for all a 2 . Clearly, these matrices Wa and Wa0 are in Y . Let L be the set of all matrices Wa and Wa0 , where a 2 . Consider now a matrix monoid S generated by fAg L. By the claim (3.4), S is mortal if and only if there is a matrix W generated by matrices of L such that AWA = 0. From the de nition (3.6), and from the properties of the morphism , it follows that W is of the form 0 3juj 0 01 W = @ 0 3jvj 0A ; (u) (v) 1 where u is in and v in ? . By (3.3) we have to analyse the condition 3juj + (u) ? (v) = 0: This is, by the property (3.1) of , equivalent to the equation v = a1 u. Because u = h(w) 2 and a1 2 , we must have = 0 W = Ww1 Ww2 Wwn where wi's are in and w = w1 wn. This is equivalent to the condition v = a1 g(w) = a1h(w): Therefore S is mortal if and only if the instance (h; g) of PCP has a solution, and this proves the claim. This proof works also for n n matrices where n > 3. We just add zero columns to the right hand side and zero rows to the bottom of all matrices in the proof of Theorem 3.1 to get n n matrices. The products of these new matrices depend only on product of original 3 3 matrices. 14

integer entries and n 3. As mentioned before, the idea of the proof of Theorem 3.1 is the same as in the original proof by Paterson. We used only one special matrix A when Paterson had two of them. Also, by the choice of this A, we managed to simplify some details in the proof. As mentioned before the PCP(7) is proved to be undecidable in MSe]. Therefore the size of the alphabet in the proof can be set to be seven and therefore we need 15 matrices in that proof.

Corollary 3.1. The mortality problem is undecidable for n n matrices with

3.2 The mortality of semigroups with two generators

In this section we consider the mortality problem in the case, where the semigroup is generated by only two matrices with integer entries. We shall show that this problem is undecidable. The proof is based on Theorem 3.1 and on a simple trick to represent a matrix semigroup with k generators of dimension n in a semigroup with only two generators of dimension nk. The proof of the following theorem is from CKa]. Theorem 3.2. Given two square matrices A and B with integer entries, it is undecidable whether the semigroup generated by fA; B g contains the zero matrix. Proof. By Theorem 3.1 we know that the presence of the zero matrix is undecidable for a semigroup T generated by M1; : : : ; Mk with dimension n = 3, where k = 15. We shall construct two matrices A and B of dimension nk such that the semigroup S generated by fA; B g contains the zero matrix if and only if T contains it. Since an algorithm to decide this property for S could then be turned into an algorithm for any nitely generated semigroup, also for T , this shows that the problem is undecidable. The construction is quite simple. A and B are de ned with n n blocks, using the matrices Mi, the n n identity I and the n n zero. A is a block diagonal and B is a permutation matrix:

0M1 0 B 0 M2 A = B .. . . B @. .
0

... ... 0 Mk

0 1 0 0 0 BI 0 ... C B C and B = B0 I B. . C B .. . . 0A @
0 15

0 I 0 0C C 0 0C C . . . . . . ... C A 0 I 0

It is quite clear that B ?1 = B k?1 = B T and that B k = Ink . We shall prove next that for any index 1 i k, the element Ci = B k?i+1AB i?1 is a block diagonal like A, but with the blocks circularly permuted, namely its diagonal is Mi; Mi+1 ; : : : ; Mk ; M1 ; : : : ; Mi?1. The proof is by induction on i. First, if i = 1, then C1 = B k?1+1AB 1?1 = B k AI = IAI = A. Assume now that there exists i such that the claim holds for all t, 1 t i < n. Then Ci+1 = B k?(i+1)+1 AB i+1?1 = B ?1CiB , and

0 1 0 0 0 Mi BMi+1 0 0 0C B C B 0 Mi+2 0 ?1 C B = B T B 0C B i B ... . . . . . . . . . ... C C @ A 0 0 Mi?1 0 0Mi+1 0 01 B 0 Mi+2 . . . ... C C = B .. B @ . ... ... 0 C : A
0 0 Mi

So Ci+1 is of the required form. Using the relation B k = Ink , any element of the semigroup S can be written in the form B tCi1 Ci2 Cim with m 0. Therefore, if we assume that the zero matrix is in S , we have Ci1 Ci2 Cim = 0, since B is invertible. Now the left upper block of the product Ci1 Ci2 Cim is 0 = Mi1 Mi2 Mim , which means that the zero matrix is in T . Conversely, assume that T contains the zero matrix and that Mi1 Mi2 Mim = 0. Let Dj = Ci1?j+1Ci2?j+1 Cim ?j+1 for 1 j k, where the indices are taken modulo k. The matrix Dj is block diagonal and its j 'th diagonal block is Mi1 Mi2 Mim = 0. Therefore, the product D1 D2 Dk , which is an element of S , is the zero matrix. It follows from the previous theorem that the mortality problem is undecidable for semigroups generated by two matrices, if the dimension of the matrices is at least 3 15 = 45. We shall return to the question of the mortality of semigroups generated by two matrices in the Chapter 7, where this problem is shown to be decidable for 2 2 matrices. 16

3.3 The existence of the zero element in matrix semigroups

In the two previous sections we considered the existence of the zero matrix in a nitely generated matrix semigroup. In this section we consider the existence of the zero element in a matrix semigroup, i.e. the existence of an element x in a semigroup S such that, for all a in S , xa = ax = x. Clearly, if a matrix semigroup S contains the zero matrix, then it has also a zero element. We shall show in this section that also the existence of the zero element is an undecidable property. We begin with an easy semigroup theoretical lemma.

Lemma 3.1. The zero element in a semigroup is unique.

Proof. Let (S; ) be a semigroup and to the contrary that S contains two zero elements, say x1 and x2 . Then by the de nition of a zero element x1 = x1 x2 = x2 .

Next theorem is obvious after the previous lemma.

Theorem 3.3. For n 3, it is undecidable whether a given nitely generated matrix semigroup of n zero element.

n matrices with integer entries contains the

Proof. Assume that the matrices M1 ; : : : ; Mk are the given generator matrices of the semigroup. Assume to the contrary that it is decidable whether S contains a zero element. Then we can also decide whether the zero matrix is in S . This follows, since if there does not exist a zero element in S , then the zero matrix is not in S either, and if there exists a zero element, we check all the elements in S until we nd it. The checking is easy to do, since X is a zero element if and only if

M1X =

= Mk X = X = XM1 =

= XMk :

Now when we nd the zero element, we check if it is the zero matrix or not. Hence, indeed, we can decide whether S contains the zero matrix, a contradiction with Theorem 3.1. 17

3.4 The mortality problem in dimension two

It is clear that for 1 1 matrices the mortality problem is decidable. The only possibility to have the zero matrix in S is that it is one of the generators. What can we say about the problem for 2 2 matrices? Not much, because it is not known whether the mortality problem is decidable or undecidable for 2 2 matrices in the case, where we have more than two generators. As we mentioned before, it can be proved decidable in two generator case. This proof is presented in Chapter 7. One thing we can say is that if the mortality problem for 2 2 matrices with arbitrary many generators is undecidable, then it must be proved with some other techniques than the undecidability for 3 3 matrices, because that proof was based on the injective morphism from into N 3 3 , and 2 2 . We shall now prove this fact, but before there is no such morphism into N that, we prove a lemma concerning the commutation of certain special 2 2 matrices.

Lemma 3.2. Let A be an upper triangular 2 2 matrix over C . b 6 1) If A = a a , where b = 0, then for all matrices B 2 C 2 2 , AB = 0
e BA if and only if B = 0 f . e 0 2) If A = a d , where a 6= d, then for all B 2 C 2 2 , AB = BA if and 0 e 0 only if B = 0 h .

e Proof. 1) Assume that B = g f and that AB = BA. These imply h that (AB )11 = ae + bg = ae = (BA)11 and therefore g = 0. Now we see that (AB )12 = af + bh = eb + fa = (BA)12 , so h = e. Clearly also (AB )21 = 0 = (BA)21 and (AB )22 = ae = (BA)22 . This shows that if B commutes with A it is of the required form, and conversily. e 2) Assume that B = g f and that AB = BA. Then (AB )12 = h af = fd = (BA)11 , and since a 6= d, it follows that f = 0. Also (AB )21 = dg = ga = (BA)21 , so g = 0 and we are done, since the other direction is obvious.
The reason we need the previous lemma is that we use in the next proof the existence of so called Jordan normal form of square matrices. The Jordan 18

normal form J of a matrix A in Zn n is an upper triangular matrix, which is similar to A, that is, there exists an invertible matrix P in C n n such that, A = PJP ?1. The diagonal of J consists of the eigenvalues of A. For a matrix A in Z2 2, the Jordan normal form has two possibilities: 1) If A has two di erent eigenvalues, say and , then J = 0 0 . 2) If A has only one eigenvalue, say , then then J = 0 0 or J= 0 1 . For the proof of these facts we refer to Gan]. The next theorem shows that there is no injective morphism from into C 2 2 . Actually we need this property only for morphisms into N 2 2 , but we use C 2 2 here, because in the proof we need the Jordan normal form of matrices, which is in C 2 2 , since the eigenvalues are in C . Now we are ready for the theorem, the proof of it being from CHK].

Theorem 3.4. There is no injective morphism : ! C2 2

for any alphabet with at least two elements. Proof. It is su cient to prove the theorem in the case where this case the monoid S = has a generating set

= f0; 1g. In

L = f(0; ); (1; ); ( ; 0); ( ; 1); ( ; )g;

where is the empty word. To simplify the notations we set a = (0; ), b = (1; ), c = ( ; 0), d = ( ; 1) and e = ( ; ). Assume to the contrary that there is an injective morphism from S into C 2 2 , and let A = (a), B = (b), C = (c), D = (d), and E = (e). Since the conjugation by an invertible matrix does not in uence the injectivity, we can suppose that A is in the Jordan normal form. Suppose that A has two di erent eigenvalues. Then A = 0 0 , and the matrices commuting with A are exactly the diagonal matrices by Lemma 3.2. Therefore C and D must be diagonal, since ac = ca and ad = da in 19

S . It follows that also matrices C and D commute, which contradicts the injectivity, since c and d do not commute in S . If A has only one eigenvalue, then A = 0 0 or A = 0 1 . The rst case is impossible, because then A = I commutes with all matrices, especially with B , but b and a do not commute in S . In the second case y A would commute only with the matrices of the form x x by Lemma 0 3.2, and they commute with each others, which yields a contradiction as above.
The mortality of 2 2 matrices is also connected to other branches of mathematics, cf. Sch]. We end this chapter by showing that the mortality problem is decidable for the 2 2 upper triangular matrices. This special case is not very important and it does not say anything about the general case.

Theorem 3.5. Let L be a set of 2 2 upper triangular matrices with integer

A11 = 0 and B22 = 0:
Proof. Let A and B be 2 2 upper triangular matrices,
11 a A = a0 a12 22 11 b and B = b0 b12 : 22

entries. Then the zero matrix belongs to the semigroup generated by L if and only if there are matrices A and B in L such that

Then

AB = a110b11 a11 b12 + a12 b22 : a22 b22

We see that the product is also an upper triangular matrix. To get zeros to the diagonal, necessarily either a11 or b11 equals with the zero and either a22 or b22 equals with the zero. So we have four cases, but, by symmetry, it is enough to consider only two of those. First, if a11 = a22 = 0, then AA = 0. The second case is that a11 = b22 = 0, and then AB = 0. Both directions of the equivalence follow from this.

20

4 Freeness of Matrix Semigroups

In this chapter we concentrate on the freeness property of semigroups which is one of the fundamental properties of semigroups and monoids. Recall that a semigroup S is said to be free if there exists a subset X of S such that every element of S has a unique factorization over X , i.e. every element of S can be uniquely expressed as a product of elements of X . The results in this chapter also concern monoids, because a monoid M is said to be free if M nf1g is a free semigroup. Next we de ne the freeness problem for square matrices.
matrices with non-negative integer entries is free?

Problem 2. Determine whether a given nitely generated semigroup of n n

As in the case of the mortality problem, we rst prove that this problem is undecidable for 3 3 matrices, and therefore also for all n n matrices, where n 3. After that we consider the case of 2 2 matrices. In the nal section of this chapter we prove that it is undecidable, whether two matrix semigroups have an equal element.

4.1 The undecidability of the freeness

We shall now prove that the freeness problem is undecidable for 3 3 matrices. This result was rst proved by Klarner, Birget and Satter eld KBS] in 1990, but we will present a proof which was developed by J. Cassaigne, T. Harju and J. Karhumaki CHK]. This proof is shorter and also gives a better bound for the number of matrices. The proof uses the same techniques than the proof of the mortality problem, but instead of an instance of PCP we will reduce an instance of MMPCP to this problem. Assume that is an alphabet, = fa1 ; : : : ; ang, and let function : ! N correspond a value of word as an n-adic representation as in the mortality chapter. 21

We now de ne a mapping 1 :
1 (u; v ) = (

0njuj 0 (u; v))T = @ 0 njvj

0 0

! N 3 3 by setting
(u) (v)A ; 1

where is the injective morphism de ned in Section 3.1. Clearly also 1 is a doubly injective morphism. Theorem 4.1. It is undecidable whether a semigroup generated by nite set of upper-triangular 3 3 matrices of non-negative integers is free. Proof. Let (h; g) be an instance of MMPCP. Of course, we may assume that h and g are morphism from into i.e. they are endomorphisms. Next de ne the set

M = f 1(a; h(a)); 1(a; g(a)) j a 2 g

and let S be the semigroup generated by M . Let i1 ; : : : ; ip ; j1 ; : : : ; jq be in M , it = 1(ait ; hit (ait )) and js = (bjs ; gjs (bjs )), with hit and gjs in fh; gg and ait and bjs in , for t = 1; : : : ; p 1 and s = 1; : : : ; q. Then, by the de nition of 1 (and ), we have:
i1 : : : ip

j1 : : : jq

in S =(
j1 : : : jq )2;3 :

if and only if (
i1 : : : ip )1;3

=(

j1 : : : jq )1;3

and (

i1 : : : ip )2;3

But this is equivalent to

ai1 : : : aip = bj1 : : : bjq and hi1 (ai1 ) : : : hip (aip ) = gj1 (bj1 ) : : : gjq (bjq )
by the uniqueness of the n-adic representations. We have proved that S is nonfree if and only if the instance (h; g) of MMPCP has a solution. Hence the nonfreeness, and so also the freeness, is an undecidable property. We could have used also the morphism itself in the above proof but we prefered to formulate the result for upper triangular matrices. This is the rst reason that this proof is better than the proof of Klarner, Birget and Satter eld KBS]. The second reason is that they needed 29 matrices when we needed only 18, assuming that in the PCP we consider instances over 7 22

letter domain alphabet, as can be done, cf. MSe] - in the undecidability proof of the MMPCP we added two letters to this alphabet. The next corollary is clear by extending the matrices in the above proof in an obvious way.
gular matrices with non-negative integer entries for any n 3 .

Corollary 4.1. The freeness problem is undecidable for n n upper trian-

4.2 The freeness problem in dimension two

As in the case of the mortality problem, it is not known whether the freeness problem is decidable or undecidable in the case of 2 2-matrices. But, again, we know that if it is undecidable, it cannot be proved using a construction similar to the one used for 3 3-matrices. This is due to Theorem 3.4. In the theory of semigroups, it is known that every nitely generated free semigroup is isomorphic to some free word semigroup + , where k is k the number of letters in the alphabet . It is also known that every free semigroup + can be embedded into the semigroup of 2 2-matrices over k nonnegative integer, i.e. into N 2 2 . Such embeddings can be done in many di erent ways. For example, if k = fa0 ; a1; : : : ; ak?1g, then the morphism de ned by

i ai 7! k 1 0
is one such embedding. Another one is the embedding de ned in the chapter considering the mortality problem, Chapter 3. It is known that it is decidable whether a nite set of words is a free generating set, i.e. whether a semigroup generated by a nite set of words is free, cf. BeP]. Therefore it is natural to ask, whether the freeness problem is decidable for the 2 2 matrices over nonnegative integers. This problem seems to be very di cult. Indeed, in CHK] Cassaigne, Harju and Karhumaki considered a restricted case, where the semigroup is generated by two upper triangular 2 2 matrices over Q , but even in this case, no general method for deciding the freeness was found. 23

4.3 Common elements in semigroups

In this section we consider the existence of a common element in two matrix semigroups. Problem 3. Given two nitely generated subsemigroups of Mn(Z), say M and N , determine whether there exists a matrix X , which is both in M and in N . Here `given' means again that we are given the generators of M and N . The existence of a common element is actually an independent problem and it is not connected to the freeness problem, but we present it here, since in the proof of the undecidability we use the same mapping 1 as de ned earlier in this chapter. M. Krom considered a variant of this problem in Kro]. He asked, whether there exists a matrix X in both M and N such that X is a product of equally many generators of both M and N . We shall prove that the problem of equal element is undecidable, if n 3, and the proof we present suits also for the variant of Krom. Theorem 4.2. It is undecidable whether two nitely generated subsemigroups of M3 (Z) have a common element. Proof. Let (h; g) be an instance of PCP, where h and g are morphism from into . Assume that and let 1 be the morphism de ned in Section 4.1. Now let SH and SG be two semigroups generated by the matrices Ha = 1(a; h(a)) and Ga = 1(a; g(a)) for all a in , respectively. Since 1 is doubly injective, we get that

Ha1
if and only if

Hak = Gb1

Gbt

a1 : : : ak = b1 : : : bt and h(a1 ) : : : h(ak ) = g(b1) : : : g(bt ):

Therefore the claim follows from the undecidability of the PCP. Clearly this proof works for the variant of Krom, since in the proof neccessarily k = t.

24

5 Finiteness of Matrix Semigroups

We are given a nitely generated subsemigroup S of N n n , i.e. we are given the generator matrices of S . In this chapter we consider the problem which asks: Problem 4. Determine whether the semigroup S is nite or not? We show that this problem is decidable by showing that, if S is nite, then an upper bound of the cardinality of S can be computed. Natural extensions of the niteness problem are the ones considering semigroups of matrices over the rational numbers Q and, in general, over any eld. Also these extensions can be proved to be decidable, cf. MaS]. Our proofs and notions are from the article MaS] by A. Mandel and I. Simon in 1977. At the end of this chapter we mention also some other results which give better upper bounds than the bound proved here.

5.1 The decidability of the niteness

As mentioned above, we shall show that there exists an upper bound for the cardinality of a nite subsemigroup of N n n , which depends on n and on the number of generators of the semigroup. The proof we present is a combinatorial one. First we need some new de nitions. Let S be a semigroup. An element a of S is called a torsion (or periodic), if ap = aq for some natural numbers p < q. If every element of S is a torsion, then S is a torsion semigroup (or a periodic semigroup). Note that a matrix A from M (N ) is torsion if and only if there exists a natural number m, such that (Ar )ij m for all r. We need also the next theorem, which is a classical result of Ramsey Ram]. Theorem 5.1. Given natural numbers m, n and k such that m 2 and n k 1, then there exist a natural number R(m; n; k) such that for every set X of cardinality at least R(m; n; k) and every partition of k element subsets of X into m blocks, there exists a subset Y of X , with the cardinality n, such that all k element subsets of Y belong to the same block. 25

Ramsey's theorem has many equivalent formulations and there exists a branch in combinatorics, namely The Ramsey Theory, which is based on this theorem, cf. GRS]. Next we de ne a graph representation of a matrix in Mn(N ). For a given A in Mn(N ), we de ne a graph GA to have a vertex set V = n = f1; : : : ; ng and an edge set E , such that, there is a directed edge from i to j in E with label Aij , whenever Aij 6= 0. For an edge e, we denote its label by (e). For a walk T = (i0; e1 ; i1; : : : ; er ; ir ) in GA, where ij 's are vertices, and ei is an Q edge from ij?1 to ij , we de ne the label (T ) to be r=1 (ej ). We say that j the length of T is r. The next lemma follows from the above de nitions. Lemma 5.1. Assume that A is in Mn(N ). For all r in N , (Ar )ij is the sum of the labels of the length r walks from i to j in GA . Proof. We prove this by induction on r. If r = 1, then this is clear by the de nition of GA. Assume that for some k 1, the claim holds, whenever r k. This means that for all i and j , (Ar )ij is the sum of the labels of all walks from i to j of length r. Now (Ak+1)ij = (A Ak )ij =
k X h=1

Aih(Ak )hj ;

which is the sum of all labels of walks of length k + 1 from i to j . Before our next lemma we de ne one more property of graphs. Let G be a directed graph, G = (V; E ), where V is the set of vertices and E is the set of directed edges. Then a subgraph H of G, H = (VH ; EH ), where VH V and EH E , is called a strong component of G, if for all i and j in VH , there is a directed path (or walk) from i to j and from j to i. Lemma 5.2. Let A be a matrix in Mn(N ). Then the following statements are equivalent: (a) A is a torsion. (b) GA contains neither a directed cycle with label at least 2, nor two directed cycles which are connected by a path. (c) There exists a permutation matrix P such that P ?1AP has the block form 0B B B 1 @ 011 B12 B13A ; (5.1) 22 23 0 0 B33 26

where some blocks might be empty, B11 and B33 are upper triangular with zeros on the diagonals, and B22 is a permutation matrix. Proof. To show that (a) implies (b), let A be in Mn (N ) and let GA be its graph representation. Assume that GA contains a cycle T with (T ) 2, and assume also that i is a vertex in T . Denote by l the length of T ; clearly l 1. By Lemma 5.1, (Alr )ii (T )r 2r , and hence A is not a torsion. Assume now that GA has two distinct cycles T1 and T2 through vertices i and j , respectively, and that there is a path T3 from i to j . Let lk be the length of Tk for k = 1; 2; 3. Consider the walks of the form T1m1 l2 T3T2m2 l1 . The length of such a walk is clearly (m1 + m2 )l1l2 + l3 . If m = m1 + m2 , then for all m in N we have at least m di erent paths from i to j of length ml1 l2 + l3, because there is m di erent ways to express m as a sum m1 + m2 . Therefore by Lemma 5.1, (Aml1 l2+l3 )ij m, and so A is not a torsion. Next we show that (b) implies (c). We say that a strong component of the graph GA is trivial if it contains no edges. Consider the following partition of the vertex set of GA:

V2 = fi 2 n j i belongs to a nontrivial strong component of GAg; V1 = fi 2 nnV2 j there exists a path from i to some vertex in V2g; V3 = nn(V1 V2 ):
Let Gk denote the subgraph of GA induced by Vk for k = 1; 2; 3, i.e. Gk consists of all vertices of Vk and all edges between the vertices of Vk . Because the vertices that are in some cycle of GA belong to V2, it follows that G1 and G3 are acyclic. Therefore we are able to de ne total orders 1 and 3 on V1 and V3, respectively, such that if there is a edge from i to j in Gk , then i k j . Now there exists a permutation of n, such that

8 i; j 2 Vk ; i 6= j : (i) < (j ) () i <k j for k = 1; 3; (5.2) 8 i 2 Vk ; j 2 Vl : k < l =) (i) < (j ): (5.3) Such a permutation can be constructed in such a way that n is rst divided

into three parts according to (5.3), and then the permutation is de ned using (5.2). Let P to be the permutation matrix de ned by Pij = i; (j). Then the matrix P ?1AP satis es the condition (P ?1AP )ij = A (i); (j) , since the multiplication by P ?1 = (i);j from the left permutes the rows of A according to , i.e. (P ?1A)ij = A (i);j , and the multiplication by P from the right permutes the columns of A (i)j according to . Let Bij denote the restriction of the matrix P ?1AP to (Vi) (Vj ). Because no vertices of V3 is connected 27

to V1 or V2 (by de nition), it follows from (5.3) that B31 and B32 are zero matrices. Also B21 is zero, because there exists no path from the vertices of V2 to the vertices of V1, only vice versa. Further since no element in V1 V3 belongs to any nontrivial strong component of GA, it follows that B11 and B33 have zeros on the diagonals: from (5.2) it also follows that they are upper triangular. Conditions (b) imply that the strong components of GA are cycles with label 1, and because no two distinct cycles are joined by a path, there are no edges, such that i and j belong to di erent strong components and there is a an edge from i to j in GA. Therefore B22 is a permutation matrix. Finally we show that (c) implies (a). Clearly, for a permutation matrix P there exists a natural number r such that P r = I . Let r be such a natural r number that B22 = I and r n. Now we have 00 B 0 B 0 1 12 13 0 (P ?1AP )r = P ?1Ar P = @0 I B23 A ; 0 0 0 because B11 and B33 are upper triangular with zeros on the diagonals; so n n B11 = B33 = 0. Hence 00 B 0 B 0 B 0 1 12 12 23 0 B23 A P (P ?1AP )2r = P ?1A2r P = P ?1 @0 I 00 B00 B0 0B0 1 0 12 12 23 0 B23 A P: (P ?1AP )3r = P ?1A3r P = P ?1 @0 I 0 0 0 Therefore A2r = A3r , i.e. A is a torsion. As corollaries of the previous lemma we prove two lemmas, but rst we must set a few new de nitions. De ne a semiring N 2 = f0; 1; 2g N with the operations a b = minfa + b; 2g and a b = minfab; 2g. Let : Mn(N ) ! Mn(N 2 ) be the monoid morphism given by (A)ij = minfAij ; 2g and let denote the set inclusion : Mn (N 2 ) ! Mn(N ). Note that, for a matrix A in Mn (N ), G (A) has exactly the same edges than GA, and if GA has a cycle with a label at least 2, then so does G (A), since in GA, necessarily one of the edges on this cycle has a label at least 2, and in G (A) this label is exactly 2. Similarly, if G (A) contains a cycle with a label 2, (the product of labels of a path is ), then GA contains a cycle with a label at least 2. Therefore, the graph GA satis es the condition (b) of Lemma 5.2 if and only if the graph G (A) satis es it. 28

We shall use these considerations as the nal conclusion in the proof of the next lemma.

Lemma 5.3. Let A and B be matrices of Mn(N ) such that (A) = (B ).

Then A is a torsion if and only if B is a torsion. Proof. Because (A) = (B ), graphs G (A) and G (B) are equal. Therefore, if GA satis es Condition (b) of Lemma 5.2, so does GB .

Recall that an element a of a semiring is called idempotent, if a2 = a. The next lemma shows how the torsions of the monoid Mn (N ) are connected to the idempotents of the monoid Mn(N 2 ). Lemma 5.4. If A 2 Mn(N ) is a torsion and (A) is an idempotent, then A2 = A3 , and there exists a permutation matrix P such that P ?1AP has a block form

00 C D1 @0 I E A ;
0 0 0

(5.4)

where I is an identity matrix. Proof. Let P be a permutation matrix such that P ?1 AP has the form (5.1) of Lemma 5.2. Since (A) is idempotent in Mn (N 2 ) and is a morphism,

((P ?1AP )2) = (P ?1A2P ) = (P ?1AP ): Now since B11 and B33 are upper triangular with zeros on the diagonals, B11 and B33 must be zero matrices. Indeed, if they are not zero matrices, then on each power the number of zero entries grows. Since B22 is a permutation 2 matrix and therefore invertible, from B22 = B22 it follows that B22 = I . Hence P ?1AP is of the form (5.4). Now

00 (P ?1AP )2 = P ?1A2 P = P ?1 @0 00 0 ?1 AP )3 = P ?1A3 P = P ?1 @0 (P

C I 0 C I 0 0

CE E A P; 0 1 CE E A P: 0

So we have that P ?1A2P = P ?1A3P , and by multiplying this equation by P from the left and by P ?1 from the right we get that A2 = A3. 29

We are ready for the main theorem of this section. It shows that if a nitely generated subsemigroup S of Mn (N ) is nite, then we can state an upper bound for cardinality of S .

Theorem 5.2. Let S be a subsemigroup of Mn (N ), generated by k of its

elements. The following statement are equivalent: (a) S is nite. (b) For all A 2 (S ), if A is idempotent then (A)2 = (A)3 . (c) jS j g(n; k) where g is a function depending only on n and k.

Proof. First we show that (a) implies (b). Since S is nite, it is torsion. Let A in (S ) be idempotent and let B in S be such that (B ) = A. Since A = ( (A)), it follows that (B ) = ( (A)). Since B is a torsion, by Lemma 5.3, so is (A), and by Lemma 5.4, since A is idempotent, (A)2 = (A)3 . Second we show thatP implies (c). Let m = 3n2 = jMn(N 2 )j, p = (b) ?1 R(m; 4; 2) and g(n; k) = p=0 ki. Without loss of generality we may assume i that S is a monoid. Since S is generated by k of its elements, there is an epimorphism, i.e. a surjective morphism, : X ! S , where X is the free word monoid generated by an alphabet X of cardinality k. We continue by proving the next claim. Claim. If x in X is a word of length jxj p, then there exists another word y in X , with jyj < jxj, such that (x) = (y). Proof of the Claim. Let r = jxj and let x = x1 x2 : : : xr , with xi 's in X . Let us partition the 2-subsets of f1; : : : ; rg into m blocks fQA j A 2 Mn(N 2 )g; by letting

QA = ffi; j g j i < j and ( (xixi+1 : : : xj )) = Ag:

Since r p, there exists, by Theorem 5.1, a set Y = fi1 ; i2; i3 ; i4g, with 1 i1 < i2 < i3 < i4 r, such that all 2-subsets of Y belong to the same block, say QA. Let yk = xik : : : xik+1 ?1 for k = 1; 2; 3, and let u and v in X be such that x = uy1y2y3v. Denote the composition ( ) by . Since fi1; i2 g, fi2 ; i3g, fi1 ; i3g and fi3 ; i4g are in QA, we have

A = (y1) = (y2) = (y1y2) = (y3):

(5.5)

Now, because is a morphism (since and are), we have that A = (y1y2) = (y1) (y2) = A2, i.e. A is an idempotent in (S ). By our assumption (b), this implies that (A)2 = (A)3 , and by Lemma 5.4, there exists a permutation matrix P such that P ?1 (A)P has a block form (5.4). Since P and P ?1 are permutation matrices, it is clear that (P ) = P and 30

(P ?1) = P ?1, and it follows from (5.5) that P ?1 (A)P = (P ?1 (yk )P ), for k = 1; 2; 3. Hence by the de nition of and , P ?1 (yk )P has the block form

00 C D 1 @0 Ik Ekk A :
0 0 0

This implies that

00 C C E 1 00 C D 1 1 1 2 3 3 (P ?1 (y1)P )(P ?1 (y2)P )(P ?1 (y3)P ) = @0 I E2 A @0 I E3 A 0 0 0 0 0 0 00 C C E 1 1 1 3 = @0 I E3 A = (P ?1 (y1)P )(P ?1 (y3)P ):

0 0 0

Now P ?1 (y1y2y3)P = P ?1 (y1y3)P , i.e. (y1y2y3) = (y1y3). Thus, (x) = (uy1y3v) and this completes the proof of the Claim, since jy2j > 0. It follows that (x) = (z) for some word z in X , such that jzj < p; hence jS j g(n; k). It is clear that (c) implies (a). In our proof we had p = R(m; 4; 2), but these Ramsey numbers are very large. We mention that A. Weber and H. Seidl proved in WeS] that the number R(m; 4; 2) can be replaced by the number de2 n!e? 1, where e is the natural base of logarithms. Their proof is based on the theory of automata, and it is di erent from the one we presented. Weber and Seidl also mention in the appendix of WeS] that it can be decided in time O(n6 jH j), whether a subsemigroup of Mn (N ) generated by the set H is nite. There exists also many other papers concerning this upper bound and the exact algorithm, for example by G. Jacob, Ja1] and Ja2], and by some of the authors already mentioned in this chapter.

Corollary 5.1. It is decidable whether a given nitely generated subsemigroup of Mn (N ) is nite. Proof. The previous theorem shows that we have to nd only g(n; k) + 1 di erent matrices from S to see that S is in nite. If we cannot nd that many di erent matrices, then S is nite.

31

Mandel and Simon also proved in MaS], that there is an upper bound for the cardinality of the nite subsemigroups of Mn(Q ) and of Mn(F ), where F is an arbitrary eld. The next two theorems are corollaries of the result of Mandel and Simon for Mn(Q ). Theorem 5.3. It is decidable whether a semigroup generated by one integer matrix is free. Proof. We can decide whether the semigroup generated by a given matrix is nite or not, and if the semigroup is nite, then it is not free, and otherwise it is free. Theorem 5.4. It is decidable whether some power of a given integer matrix is zero, i.e. the mortality problem for the semigroup generated by one matrix is decidable. Proof. We can decide whether the matrix semigroup generated by this matrix is nite or not, and if it is, we check if one of these matrices is zero. The same conclusion holds for the decidability of the existence of the identity matrix as a power of given matrix. It is also decidable.

32

6 Zero in the Right Upper Corner

The problem considered in this chapter is called the zero in the right upper corner.
an M in the semigroup generated by these matrices such that M1n = 0.

Problem 5. For a given nite subset of Zn n, determine whether there exists

In the rst section we shall show that if n 3, then the problem is undecidable. The method we use is the usual coding of pairs of words into 3 3 matrices. First we have to recall and introduce few notions and functions. In the latter part of this chapter we consider the above problem in the case where we have only two generators. We shall prove that also in this case the problem is undecidable, when the dimension of the matrices is large enough.

6.1 The undecidability of the zero in the right upper corner

We begin with some de nitions for the proof. Let ? = fa1 ; a2; :::; ang be an alphabet and let function be as in Chapter 2, i.e. : ? ! N such that (ai1 ai2 :::aik ) =
k X j =1

ij nk?j and (1) = 0:

We recall that is injective, and that for all u; v in ? (uv) = (v) + njvj (u): (6.1) Assume now that n = 2, i.e. ? = fa1 ; a2g, and de ne a mapping 2 : ? ? ! Z3 3 by

01 2 (u; v ) = @0

(v) 2jvj 0 0 33

(u) ? (v) 2juj ? 2jvj A : 2juj is injective, and it is also a

This mapping is clearly injective, because

morphism, since for all u1; u2; v1; v2 in ? , 2 (u1 ; v1 ) 2 (u2; v2 ) 01 (v ) (u ) ? (v )1 01 (v ) (u ) ? (v )1 1 1 1 2 2 2 = @0 2jv1j 2ju1j ? 2jv1j A @0 2jv2j 2ju2j ? 2jv2j A 0 0 2ju1j 0 0 2ju2j 01 (v ) + (v )2jv1j (u ) ? (v ) + (v )(2ju2j ? 2jv2j) 1 2 1 2 2 1 ju2 j ( (u1 ) ? (v1 )) B C +2 = B0 jv1 j 2jv2 j jv1 j (2ju2 j ? 2jv2 j ) + 2ju2 j (2ju1 j ? 2jv1 j )C @ A 2 2 ju1 j2ju2 j 0 01 (v v 0 (u u ) ? (v v )1 2 1 2) 1 2 1 2 (6.1) @ 2ju1u2j ? 2jv1v2 j A = 2(u1u2; v1v2 ): = 0 2jv1v2 j 0 0 2ju1u2 j The next theorem is attributed to R.W. Floyd in Man]. Theorem 6.1. It is undecidable whether a nitely generated subsemigroup of M3 (Z) contains a matrix M with M13 = 0. Proof. Let (h; g) be an instance of PCP, let h and g be morphisms from into ? , and de ne the matrices Ma = 2(h(a); g(a)) for all a in . Then, because 2 is morphism, for a matrix M = Ma1 Ma2 Mam , M13 = 0 if and only if (h(a1)h(a2 ):::h(am )) = (g(a1)g(a2):::g(am)); i.e. (h(w)) = (g(w)), where w = a1 a2 : : : am . Now, because is injective, we get that M13 = 0 if and only if h(w) = g(w), and the claim follows from the undecidability of PCP. Note that in the proof we needed 7 generators, since, as remarked before, PCP(7) is undecidable. Because we may add zero columns to the left and zero rows to the bottom of the matrices in the previous proof, we get the next corollary. Corollary 6.1. If n 3, it is undecidable whether a nitely generated subsemigroup of Mn (Z) contains a matrix M with M1n = 0. The decidability of the existence of the zero in the right upper corner for 2 2 matrices is an open question. We note again, that if it is undecidable, it must be proved with some other method than the one used in the proof of the previous theorem. 34

6.2 The two generator case

In Chapter 3 we considered the mortality problem in the case, where the semigroup is generated by two matrices. We used the construction from matrices M1 ; : : : ; Mk with dimension n into two matrices A and B with dimension nk such that 00 0 1 0 M1 0 0 I 01 BI 0 0 0C B C B 0 M2 . . . ... C C and B = B0 I 0 A = B .. . . . . B . . . . 0C C B B .. . . . . . . ... C C @. . . 0A @ A 0 0 Mn 0 0 I 0 We cannot apply these matrices to the problem of the zero in the right upper corner, but we shall use them as blocks in the matrices with dimension nk +3. First we de ne two vectors of dimension n, namely

X = 1 0 ::: 0

01 B0C .. and Y = B . C ; B0C @A

1

so XMY is the entry in the right upper corner of M in Zn n. Next de ne two vectors of dimension nk,

0 1 Y ?X : : : X and V = B 0 C : B.C U= B .. C @ A
0

Now we de ne the matrices A0 and B 0 with dimension nk + 3 to be 00 1 U 1 1 00 1 U 1 1 B B 0 1C 0 1C A0 = B0 0 A V C and B 0 = B0 1 B V C : A A @0 0 @0 0 0 0 0 0 0 0 0 0 Note the di erence in the second diagonal entry. The proof of the next theorem is from CKa]. Theorem 6.2. It is undecidable whether the semigroup generated by fA0; B 0g has an element such that it has a zero in the right upper corner. 35

Proof. Let S be the semigroup generated by fA; B g, S 0 be the semigroup generated by fA0; B 0g, and T be the semigroup generated by fM1 ; : : : ; Mk g. Let C 0 be an element of S 0. Then 1 00 B 0 C C 0 = B0 0 C C ; A @0 0 0 0 0 where 's represent the unimportant values, C is an element of S , and is 1, if C 0 is power of B 0, and 0 otherwise. Consider any element of S 0 other than the generators A0 and B 0. It can be written in the form PC 0Q, where P and Q are equal to A0 or B 0 , 00 1 U 1 1 00 1 U 1 1 B B 1C 1C P = B0 0 0 V C and Q = B0 0 0 V C ; A @0 A @0 0 0 0 0 0 0 0 0 and C 0 is an element of S 0. The product expands to 1 00 + UCV C; B C PC 0Q = B0 0 0 A @0 0 0 0 0 where the right upper corner entry is + UCV . If C 0 is a power of B 0 , then C is a power of B and = 1 and UCV = 0, hence the right upper entry of PC 0Q is 1. Otherwise, = 0P if we divide C to n n blocks Nij , and where 1 i; j k, then UCV = k=1 XNi1Y . As in the proof of Theorem i 3.2, we can write C = B t Ci1 Ci2 Cim , where Cij = B k?ij +1 AB ij ?1, where 1 ij k. The initial B t permutes the lines only, so the Ni1 's are also, in a di erent order, the elements of the rst block column of Ci1 Ci2 Cim , and one of them is M = Mi1 Mi2 Mim and the k ? 1 others are zero blocks. Therefore the right upper corner entry of PC 0Q is XMY , i.e. the right upper corner entry of M , which can be any element of T . If we now de ne the matrices Mi to be the matrices of the proof of Theorem 6.1, we get that for S 0 the existence of the zero in the right upper corner is undecidable, since it is undecidable for T .

Notice that dimension of A0 and B 0 is nk +3 and, by the proof of Theorem 6.1, we may set n = 3 and k = 7. Therefore we get that the problem of the zero in the right upper corner is undecidable in two generator semigroups, when the dimension of these matrices is at least 24. This is an obvious corollary of the previous theorem. 36

7 Skolem's Problem
In this and next chapters we consider the so called Skolem's problem, which we de ne next. It is related to the problem of the zero in the right upper corner in a way that in Skolem's problem we consider semigroups generated by one matrix. Problem 6. Let M be a given matrix from Mn(Z). Determine whether there exist a power k such that (M k )1n = 0? It is not known whether the Skolem's problem is decidable or not, when n 3. We shall prove the decidability for case of dimension two in the rst section and the main result of the next chapter is that it is decidable, whether there exists in nitely many powers k such that (M k )1n = 0. In the proof of the decidability of the Skolem's problem in dimension two we shall use Theorem 7.1 which suits for deciding the existence of the zero in any entry of the powers of a given 2 2 matrix, and the decidability of the Skolem's problem is a corollary of that theorem. That theorem gives us also a way to prove the decidability of the mortality problem in the case, when the semigroup is generated by two 2 2 matrices. This proof is presented in the second section of this chapter.

7.1 Skolem's problem in dimension two

We are now going to prove that the Skolem's problem is decidable in the case of 2 2 matrices. To prove this we need certain properties of sequences. Let (un)1 be an integer sequence, de ned by constants u0 = c0, u1 = c1 n=1 and a formula un + a1 un?1 + a2 un?2 = 0; (7.1) where ai's are in Q . Let and are the roots of the equation y2 + a1 y + a2 = 0: (7.2) The next lemma has an elementary combinatorial proof, which we give for the sake of completeness. 37

Lemma 7.1. Let (un)1=0 be as above. Then we have two cases, namely: n (i) If = , then there exists constants a and b such that 6
for all n in N . (ii) If = , then there exists constants c and d such that un = (cn + d) n Proof. (i): If the claim holds, then by the equalities a + b = c0 and a + b = c1, we have

un = a n + b

We assume that a and b are as above and prove the claim (i) by induction. Clearly, when n = 0; 1 the claim holds. Assume that it holds for all n k + 1, where k 0. Then by the assumption (7.1) uk+2 = ?a1 uk+1 ? a2uk = ?a1 (a k+1 + b k+1) ? a2 (a k + b k ) = ?a k (a1 + a2 ) ? b k (a1 + a2 ) = a k+2 + b k+1; where the last equation follows from the fact that and are roots of the equation (7.2), and therefore ? 2 = a1 + a2 and ? 2 = a1 + a2 . (ii): Like in the previous case, if the claim holds, then we have that d = c0 and c = (c1 ? c0 )= . We prove the claim (ii) by induction. If n = 0; 1, then the claim holds. Assume that claim holds for all n k +1, where k 0. Then by the assumption (7.1) uk+2 = ?a1 (c(k + 1) + d) k+1 ? a2 (ck + d) k = k (?a1 c(k + 1) ? a1d ? a2 ck ? a2 d) = k (ck(?a1 ? a2) + d(?a1 ? a2) ? a1 c ) = k (ck 2 + d 2 ? (?2 )c ) = (c(k + 2) + d) k+2; since is a root of equation (7.2) and t2 + a1 t + a2 = (t ? )2, we have a1 = ?2 . Let p be a prime number. We de ne the p-adic valuation vp over Q by vp(0) = 1 and vp(q) = n, where q = pn a with a and b in Z and p divides b neither a nor b. It is clear by the de nition of vp that vp(a + b) inf fvp(a); vp(b)g. And if a and b are integers, we can prove a stronger result for vp(a + b) as stated in the next lemma. 38

a = c1 ? c0 and b = c1 ? c0 : ? ?

vp(b), then vp(a + b) = vp(a). Proof. Assume that vp (a) = n and vp(b) = k, with n < k. Then a + b = pn(a1 + b1 ), where vp(a1) = 0 and vp(b1 ) > 1. Clearly vp(a + b) n, and if vp(a + b) > n, then a1 + b1 0 (mod p) and therefore a1 ?b1 0 (mod p), and we have a contradiction. Now vp(a + b) = n = vp(a).
Now we are ready for the main theorem of this section. The proof is modelled after the ideas of J.Cassaigne Cas]. Theorem 7.1. Let A be a 2 2 matrix over integers and assume that u and v are integer vectors of dimension 2. Then it is decidable whether uAnvT = 0 for some n > 0. Proof. Let t = tr(A) and d = det(A). By Section 2.3 we know that the characteristic polynomial of A is cA( ) = 2 ? t + d. Also by Section 2.3 we know that cA(A) = A2 ? tA + dI = 0, so we see that

Lemma 7.2. Let a and b be two integers and p a prime number. If vp(a) <

A2 = tA ? dI: (7.3) Consider the sequence (xn)n 0 = uAnvT . By multiplying the equation (7.3) by uAn from the left and by vT from the right we get xn+2 = txn+1 ? dxn. We see that xn is de ned by x0 and x1 , and if both of these are zero, then for all n, xn = 0. We assume that this is not the case. Consider next the equation (7.2) for the sequence xn, i.e. the equation y2 ? ty + d = 0: (7.4)
We know that the roots of this equation are of the form y = t t22 ?4d . Let = t2 ? 4d. We divide the proof into three cases: t 1) = 0. Then by the case (ii) of Lemma 7.1, xn = ( 2 )n(cn + d), where c and d are in Q , and they are xed by x0 and x1 . Now if t = 0,then (xn )n 0 has in nitely many zeros, and if t 6= 0 and ? d in N , then (xn)n 0 has exactly c one zero. Otherwise (xn)n 0 has no zeros. 2) > 0. If t = 0, then
2 (?d) n x0 ; if n is even; xn = b n c x1 ; if n is odd: (?d) 2

(7.5)

If x0 or x1 is zero, then (xn)n 0 has in nitely many zeros. Otherwise it has no zeros. 39

If t 6= 0, then we have case (i) in Lemma 7.1, i.e. xn = a n + b n; where and (in R ) are the roots of the equation (7.4) and a and b are in = fu + v j u; v 2 Q g R by the proof of Lemma 7.1. They can be e ectively computed by the same b proof. Assume that > . Now xn = 0 if and only if ( )n = ? a . As j j > 1, a bound on n can be found and we are able to check whether (xn)n 0 has a zero or not. 3) < 0. If t = 0, we have the case (7.5), and it is easy to verify whether xn = 0 for some n. We assume that t 6= 0. We know that d > 0 and we shall 2 prove rst that xn = O(d n ), which means that there exist constants c and 2 n0 such that for all n n0 , xn c d n . Since < 0, it follows by the case (i) of Lemma 7.1 that, for some constants a and b,
Q

t + t2 ? 4d n + b t ? t2 ? 4d n: xn = a 2 2 Now we get that p p t + t2 ? 4d n + b t ? t2 ? 4d n jxnj = a 2 p22 p2 n n a t + t ? 4d + b t ? t ? 4d 2 2 jaj jt + pt2 ? 4djn + jbj jt ? pt2 ? 4djn 2n 2n p p = j2anj jt + i 4d ? t2 jn + j2bnj jt ? i 4d ? t2jn p p n n = j2anj t2 + 4d ? t2 + j2bnj t2 + 4d ? t2 p n p n 2 2 2 = janj 4d + jbnj 4d = jajd n + jbjd n = d n (jaj + jbj); 2 2 2 2 and therefore xn (jaj + jbj)d n O(d n ). Assume that p is a prime number such that p divides t and p2 divides d. t Then pn?1 divides xn , and by setting t0 = p and d0 = pd2 , we can de ne an integer sequence (x0n )n 0 such that x00 = x0 , x01 = x1 and for n 2, n x0n = t0 x0n1 ? d0x0n?2 = px?1 : n
40

Clearly x0n = 0 if and only if xn = 0 and therefore we may consider the sequence (x0n)n 0 instead of (xn )n 0. So we may assume that there is no such prime p that p divides t and p2 divides d. Let u be the greatest common divisor of t and d, we denote, as usual, u = gcd(t; d) > 0, and t = uv, d = uw for some integers v and w. By the above assumption about prime factors of t and d, gcd(u; w) = gcd(v; w) = 1. Assume that p is a prime factor of u. Then, again by induction, we can prove that vp(xn ) n ]. Also vp(x2 ) inf fvp(tx1 ); vp(dx0 )g 1 and the 2 induction step is vp(xn+1 ) inf fvp(uvxn); vp(uwxn?1)g 1 + n?1 ] = n+1 ]: 2 2 n] This means that u 2 divides xn for all n 2. Assume now that for all prime factors p of w, and for all n, vp(xn+1) ? vp(xn) vp(w). This implies that xn 6= 0 for all n. Also since vp(w) n1, it implies that wn divides xn. Now, since gcd(u; w) = 1, we know that u 2 ]wn divides xn . Therefore 2 2 u n ]wn xn = O(d n ); and because d = uw > 0, necessarily w = 1. Now = (uv)2 ? 4u < 0, so we get that uv2 < 4, and jvj = 1 and u 2 f1; 2; 3g. Since d = u, we see that d 2 f1; 2; 3g, and t = d or t = ?d. All these cases lead to an equation xn+12 = d6xn, which can be veri ed by simple calculations. For example, if d = t = 3, then xn+2 = 3xn+1 ? 3xn. We prove the equation by induction. First, for n = 0 x12 = 3x11 ? 3x10 = 9x10 ? 9x9 ? 3x10 = 6x10 ? 9x9 = 18x9 ? 18x8 ? 9x9 = 9x9 ? 18x8 = 9x8 ? 27x7 = ?27x6 = ?81x5 + 81x4 = ?162x4 + 243x3 = ?243x3 + 486x2 = ?243x2 + 729x1 = 729x0 = 36 x0; and similarly for n = 1, we get that x13 = 36x1 . The induction step is for n 2 x(n+1)+12 = 3xn+12 ? 3x(n?1)+12 = 3 36xn ? 3 36xn?1 = 36(3xn ? 3xn?1) = 36xn+1: This proves the case d = t = 3, and the other cases can be proved similarly. We can easily verify whether xn+12 = d6xn and t = d. If this is not the case, then for some prime factor p of w and some n, vp(xn+1)?vp (xn ) < vp(w). Now vp(xn+1 ) < vp(xn) + vp(w) = vp(uwxn), so by Lemma 7.2 vp(xn+2) = vp(uvxn+1 + uwxn) = vp(xn+1 ): 41

By induction we prove that

vp(xm ) = vp(xn+1 ) < 1 for all m n + 1:

Assume that for some m, (7.6) holds for all k, m k n + 1

(7.6)

vp(xm+1 ) = vp(uvxm + uwxm?1 ) = inf fvp(xn+1 ); 1 + vp(xn+1 )g = vp(xn+1):

Clearly the only possible zero is xn where vp(xn+1 ) ? vp(xn ) < vp(w) for the rst time. So we check whether this xn = 0. Note that p and n are not known at rst, but we can compute the values vp(xn+1) ? vp(xn) for all prime factors p of w until p and n are found. There is no explicit bound for this loop of computation; it terminates always, since we evaluate this loop for all prime factors p of w at the same time. The loop terminates, when the rst pair p and n is found and then the only possibility for the zero in (xn )n 0 is found. We check that one, and we are done. We have proved that there is a method in all cases of to solve whether there is n 0 such that uAnv = 0. Corollary 7.1. Skolem's problem is decidable for 2 2 matrices over integers. Proof. If we set u = (1; 0) and v = (0; 1) in Theorem 7.1, then the product uAnvT equals (An)22.

7.2 Skolem's problem and the mortality problem

We shall now return to the mortality problem. We restrict to the case when we are given two 2 2 matrices over integers. We shall prove that this case is decidable. The proof is based on Theorem 7.1. First we consider some properties of 2 2 matrices over integers. Let A and B be elements of Zn n. We recall that det(AB ) = det(A) det(B ) and that if A is singular, i.e. det(A) = 0, we know that the rows and the columns of A are linearly dependent vectors. Set B = fv1; : : : ; vk g of vectors from Zn is linearly dependent if the exists integers c1; : : : ; ck such that

c1 v1 = c2v2 + c3v3 +
42

+ c k vk

and at least for one i, ci 6= 0. Otherwise B is called linearly independent.

The number of linearly independent rows in A is called the rank of A and denoted by rank(A). It can be proved that the number of linearly independent columns is is also rank(A). Note that the rank of a matrix is also the dimension of the vectors space generated by the rows or the columns of the matrix. Let A be inZn n. It is known that A can be presented in the form

A=

rank(A) X
i=1

aT bi; i

where ai and bi are vectors in Zn for i = 1; : : : ; rank(A). Clearly, if rank(A) = 1, then there exists vectors a and b in Zn such that A = aT b. Lemma 7.3. Let A, B and C three matrices with rank 1 from Zn n. If ABC = 0, then either AB or BC is zero. Proof. Assume that A = aT b, B = cT d and C = eT f . If one of these matrices is zero, then the claim obviously holds, so we may assume that this is not the case. It follows that ABC = aT bcT deT f = (bcT )(deT )aT f; where bcT and deT is are integers. If now ABC = 0, then clearly either bcT or deT is zero. Since AB = aT bcT d = (bcT )aT d and

BC = cT deT f = (deT )cT f;

the claim follows. Theorem 7.2. Given two 2 2 matrices over integers, say A and B , it is decidable whether the zero matrix is in the semigroup generated by fA; B g. Proof. Assume that A and B are non-zero, otherwise we are done. As we mentioned earlier, det(AB ) = det(A) det(B ) and det(0) = 0, so if A and B are both invertible, then the zero matrix is not in the semigroup S generated by fA; B g. It also follows that if the zero is in S , then at least one of these matrices is singular, i.e. has determinant equal to zero. Assume that either A or B is singular. Then by Lemma 7.3, in the minimal-length zero product of S , the singular element can occur only as the rst and the last factor. We have two possible cases: 43

1) If A and B are both singular, then we have to check only the products AB , BA, B 2 and A2 . 2) If A is invertible and B is singular, then B can be written in form B = xT y, where x and y are vectors in Z2. Now

BAk B = xT yAk xT y = (yAk xT )xT y = 0

if and only if

yAk xT = 0; and by Theorem 7.1 this is decidable.

We shall next show that the decidability status of the Skolem's problem is equivalent with the decidability status of the certain instance of the mortality problem. The decidability status of a problem is either decidable, undecidable or unknown. The theorem in this section is based on Lemma 7.3 and on next lemma. Lemma 7.4. Let A be an element of Zn n, u and v be in Zn. There exists matrix M in Z(n+2) (n+2) such that for all integers k 1

uAk vT = (1; 0; : : : ; 0)M k (0; : : : ; 0; 1)T : (7.7) Proof. De ne M as block form 00 uA2 uAvT 1 M = @0 A vT A : 0 0 0 The right hand side of the equation (7.7) clearly is the right upper corner element of M k . We shall prove by induction that 00 uAk+1 uAk vT 1 (7.8) M k = @0 Ak Ak?1vT A 0 0 0 for all integers k 1. First, for k = 1 (7.8) is obvious by the de nition. Assume now that (7.8) holds for all k j , j > 1. So 00 uAj+1 uAj vT 1 00 uAj+2 uAj+1vT 1 M j+1 = M j M = @0 Aj Aj?1vT A M = @0 Aj+1 Aj vT A : 0 0 0 0 0 0 This proves the lemma.
44

Now we are ready for the next theorem. Theorem 7.3. Let I be an instance of the mortality problem such that we are given matrices M1 ; : : : ; Mt and A in Zn n and rank(Mi ) = 1 for all i = 1; : : : ; t. The decidability status of instance I is equivalent with decidability status of Skolem's problem for matrix with dimension n + 2. Proof. By Lemma 7.3 in I it is enough to study products Mi Mj and Mi Ak Mj , i; j = 1; : : : ; t and k 1. The products of the second form are actually of the form uAk vT , since rank Mi = 1 for all i = 1; : : : ; t, and by Lemma 7.4 these are equivalent with the Skolem's problem for matrices with dimension n + 2. We have proved that there is a certain relation between the mortality problem and the Skolem's problem. Note that the decidability status of instances I of the mortality problem is unknown.

7.3 Skolem's problem for matrices over natural numbers

If we restrict in Skolem's problem to square matrices over natural numbers, we have a simple decidable special case. It follows from the fact that if M is in Mn (N ) , then in M k we are only interested in whether (M k )1n is positive or zero. Let B be the boolean semiring, B = f0; 1g, where the operations are the usual product and addition of integers except that 1 + 1 = 1. Assume that A is in Mn(N ) and de ne a mapping : Mn (N ) ! Mn (B ) by (A)ij = 0; if Aij = 0; 1; otherwise: We end this chapter with a simple theorem. Theorem 7.4. Let A be in Mn (N ). It is decidable whether (Ak )1n = 0 for some k 1. Proof. Assume that B is in Mn (B ) and that B = (A). Clearly (B k )1n = 0 if and only if (Ak )1n. There are 2n2 elements in Mn (B ), and we can compute B k for all k, and stop if (B k )1n = 0 or if B k = B h for some h < k. This procedure terminates, since the number of elements in Mn (B ) is nite. 45

8 An In nite Number of Zeros in Skolem's Problem

In this chapter we will show that it is decidable for a given matrix M from Mn(Z) whether there exist in nitely many k's such that (M k )1n = 0. The proof uses the theory of rational series so we begin with the basics of these series. Our presentation on rational series follows the book Rational Series and Their Languages by J. Berstel and C. Reutenauer BeR].

8.1 Rational series

Let X be a nite, nonempty alphabet and denote the empty word by . Recall that X is a monoid, the product is the concatenation of two words and with neutral element . Let K be a semiring and X an alphabet. A formal series (or formal power series) S is a function

X ! K: The image of word w under S is denoted by (S; w) and is called the coe cient of w in S . The image of X under S is denoted by =(S ). The support of S is the language supp(S ) = fw 2 X j (S; w) 6= 0g: The set of formal series over X with coe cients in K is denoted by K hhX ii.
Let S and T be two formal series in K hhX ii. Then their sum is given by (S + T; w) = (S; w) + (T; w) and their product by (ST; w) =

X
uv=w

(S; u)(T; v):

46

Note that here the sum is always nite. We note also that K hhX ii is a semiring under these operations. The identity element with respect to the product is the series 1 de ned by (1; w) = 1K if w = ; 0K otherwise; and the identity element with respect to the sum is the series 0 de ned by (0; w) = 0K . We also de ne two external operations of K in K hhX ii. Assume that a is in K and S in K hhX ii, then the series aS and Sa are de ned by (aS; w) = a(S; w) and (Sa; w) = (S; w)a: A formal series with a nite support is called a polynomial. The set of polynomials is denoted by K hX i. If X = fxg, i.e. X is a unary alphabet, then we have the usual sets of formal power series K hhxii = K x]] and of polynomials K x] P A formal series S can also be written in the sum form S = aw w, where w is in X and aw is the coe cient of w in K , i.e. (S; w) = aw . Example 1. Let X = fx; y; zg and consider the formal series in ZhhX ii de ned by S = xy + xxyz + 2zx. We get, for example, that (S; xy) = 1, (S; zx) = 2 and (S; xx) = 0. A family of formal series, say (Si)i2I is called locally nite, if, for every word w in X , there exists only a nite number of indicies i in I such that (Si; w) 6= 0. A formal series S in K hhX ii is called proper if the coe cient of the empty word vanishes, that is (S; ) = 0. Let S be proper formal series. Then the family (S n)n 0 is locally nite, since for any word w, if jwj > n, then (S n; w) = 0. The sum of this family is denoted by S ,

S =
n 1

X
n 0

Sn

and star of . Similarly we de P itSis .called the 0the 1, and itScan be proved that ne the plus of S , S + = n Clearly S = the equations

S = 1 + S + and S + = SS = S S
47

hold for proper formal series. The rational operations in K hhX ii are the sum, the product, the two external products of K in K hhX ii and the star operation. The subset of K hhX ii is called rationally closed, if it is closed under the rational operations. The smallest rationally closed subset of K hhX ii containing a subset E of K hhX ii is called the rational closure of E . A formal series is called rational if it is an element of the rational closure of K hX i, i.e. it can be de ned using the polynomials K hX i and the rational operations. ^ ^ Example 2. Denote by X the formal series Px2X x. X is proper, since ^ (X; 1) = 0, and rational, since it is a polynomial. The series

X^ ^ X = Xn =
n 0

w2X

w:

is also rational. We denote K m n , as usual, the set of the m n matrices over K . A formal series S 2 K hhX ii is called recognizable if there exists an integer n 1, and a morphism of monoids : X ! K n n, into the multiplicative structure of K n n, and two matrices (or vectors) 2 K n 1 and 2 K 1 n such that for all words w, (S; w) = (w) : The triple ( ; ; ) is called a linear representation of S with dimension n. A linear representation of a rational series S is called reduced, if its dimension is minimum among all linear representations of S . Note that if we de ne and as vectors in K n, then (S; w) = (w) T :

The next theorem is fundamental in the theory of rational series. It was rst proved by Kleene in 1956 for languages that are those series with coe cients in the Boolean semiring. It was later extented by Schutzenberger Scb] to arbitrary semirings.

Theorem 8.1. A formal series is recognizable if and only if it is rational.

48

The proof by Schutzenberger uses some properties of K -modules and it is quite long, and omitted here. It is clear by the de nition of recognizable series that we can transform our problem of matrices into rational series quite nicely. From now on we shall x the alphabet to be X = fxg. We denote the formal series S over X by X X S = anxn = anxn :
n 0

A rational series is called regular, if it has a linear representations ( ; ; ) such that (x) is an invertible matrix. There are also other ways to de ne regular series. It can be proved that for a regular series S , any reduced linear representation ( ; ; ) has invertible (x). Assume that K = Z and that S is a rational series in Z x]] having a linear representation ( ; ; ), and moreover that the characteristic polynomial of the matrix (x) in Zk k is c( ). Then c( ) = (?1)k ( k + c1 k?1 + + ck ); where ci's are in Z and k is the dimension of the linear representation ( ; ; ). By the Cayley{Hamilton theorem we know that c( (x)) = (xk ) + c1 (xk?1) + + ck I = 0: If we multiply this equation from the left by (xn ) (n 2 N ) and from the right by , we get an+k + c1an+k?1 + + ck an = 0: (8.1) The equation an+h + 1an+h?1 + + han = 0; P for n 0, satis ed by rational series S = anxn in Z x]] is called a linear recurrence relation. A linear recurrence relation is called proper, if h 6= 0. Note that having a linear recurrence relation, every an for n h can be computed if we know the elements a0 , a1, : : : , ah?1 of the series S . These h rst elements are called the starting conditions. Next lemma gives a method to calculate a regular linear representation from a proper linear recurrence relation. 49

relation, then it is has a regular linear representation ( 1 ; 1; 1 ), i.e. a linear representation satisfying det( 1 (x)) 6= 0. Proof. Let

Lemma 8.1. If rational series S 2 Z x]] satis es a proper linear recurrence

an+h ? 1 an+h?1 ? ? han = 0 be a proper linear recurrence relation of S and assume that aj 's, 0 j < h, are the starting conditions. Let 00 1 0 : 01 0 1 C B0 0 1 B a0 C B . . . . 0C B a1 C B. .. C 1 = (1; 0; : : : ; 0); 1 (x) = B . . . . . . . . C ; 1 = B . C : B 0 :::::::::::: 1C @ .. A A @ ah?1 h :::::::::::: 1
The matrix 1(x) is called the companion matrix. It is easy to see that det( 1(x)) = h 6= 0. Now 1 1(xj ) gives the rst row of matrix 1(xj ). If 0 j h ? 1, then the rst row of 1(xj ) has 1 in the j 'th entry and all the others are 0. Therefore 1 1(xj ) 1 = aj , if 0 j h ? 1. To prove that an = 1 (xn ) 1 for all n 0, we use the method mentioned before the lemma, i.e. we calculate the characteristic polynomial of 1(x) to get a linear recurrence relation. So we calculate ? 1 0 0 0 ? 1 0 ... . . . . . . . . . ... det( 1(x) ? I ) = ... . . . 0 ? 1 :::::::::::: 1 ? h =
h X i=1

(?1)i?1 i(? )h?i + (? )h;

where the last equality follows from the de nition of the determinant in Section 2.3: det((Aij )n n) =

sign(j1 ; j2; : : : ; jn)A1j1 A2j2

Anjn ;

where repsesents a permutation of the set f1; : : : ; ng, and from the fact that only permutations with non-zero product are of the form (1; : : : ; i ? 1; i + 50

1; : : : ; h; i). These permutations have i ? 1 inversions, and the signs of these permutations are (?1)i?1. The last term (? )h follows from the permutation (h; 1; : : : ; h ? 1) and from the element 1 ? in the product. Now det( 1(x) ?

I ) = (?1)h
= (?1)h(

h X i=1

(?1)i?1?i
1
h?1 ?

h?1 + h

Thus the rational series with linear representation ( 1; 1; 1 ) satis es the same linear recurrence relation than S . Therefore they are equal. Note that one way to de ne the regular rational series is to use the fact that they satisfy a proper linear recurrence relation.

? h):

8.2 Skolem's theorem

In this section we prove an important theorem concerning the zeros in rational series over fxg with coe cients in Z, namely a theorem proved by Skolem in 1934. Theorem 8.2. Let S = P anxn be a rational series with coe cients in Z. Then the set fn 2 N j an = 0g is a union of a nite set and of a nite number of arithmetic progressions. The proof we give here is based on the proof by G. Hansel in Han], cf. also BeR]. Hansel proved the theorem for rational series with coe cients in the elds of characteristic 0. His rst step was to prove the claim for regular rational series with coe cients in Z. We use that proof up to that point, and then present a proof for all rational series with coe cients in Z. The proof requires several de nitions and lemmas. A set A of natural numbers is called purely periodic, if there exists a natural number N and integers k1; k2; : : : ; kr 2 f0; 1; : : : ; N ? 1g such that A = fki + nN j n 2 N ; 1 i rg: The integer N is called a period of A. A quasiperiodic set of period N is a subset of N that is a union of a nite set and a purely periodic set of period N. 51

is a quasiperiodic set of period N .

Lemma 8.2. The intersection of a family of quasiperiodic sets of period N

Proof. Let (Ai )i2I be a family of quasiperiodic sets, all having period N . For all j 2 f0; 1; : : : ; N ? 1g and for all i in I , the set (j + nN ) \ Ai is either nite or equal to (j + nN ). Clearly the same holds for (j + nN ) \ (\i2I Ai ) and therefore the intersection \Ai is quasiperiodic.

Assume that S = anxn is a rational series. We denote the set fn 2 N j an = 0g by ann(S ), and it is called the annihilator of S . Clearly the annihilator is the complement of the support de ned in the previous section. Let vp be the p-adic valuation de ned in the previous chapter. Our next Lemma states an inequality, which we shall use later. Recall rst that

vp(q1
and

qn ) =

n X i=1

vp(qi )

vp(q1 +
where qi's are in Q .

+ qn ) inf fvp(q1 ); : : : ; vp(qn)g;

are dividable by p2 and so on, we have that

Lemma 8.3. If p is a prime number and n a natural number, then n vp p ! n p ? 2 : n p?1 Proof. First, since bn=pc of the numbers 1; : : : ; n are dividable by p, bn=p2c
vp(n!) bn=pc + bn=p2c + : : : + bn=pk c + : : : n=p + n=p2 + : : : + n=pk + : : : n=(p ? 1):
Therefore,
n n p?2 vp p ! = n ? vp(n!) n ? p ? 1 = n p ? 1 : n

52

Consider next an arbitrary polynomial P (x) = a0 + a1 x + : : : + anxn in Q x]. For any integer k 0, let !k (P ) = inf fvp(aj ) j j kg: Clearly, !0(P ) !1(P ) : : : !k (P ) : : : and for k > n,

!k (P ) = 1: Note also that vp(P (t)) inf fvp(a0 ); vp(a1 t); : : : ; vp(antn)g for any integer t, and therefore vp(P (t)) !0(P ): (8.2) Lemma 8.4. Let P and Q be two polynomials with rational coe cients such
that

P (x) = (x ? t)Q(x) for some t 2 Z. Then for all k 2 N !k+1(P ) !k (Q): Proof. Assume that Q(x) = a0 + a1 x + : : : + anxn ; P (x) = b0 + b1 x + : : : + bn+1xn+1 : Then bj+1 = aj ? taj+1, for 0 j n ? 1, and bn+1 = an, and therefore for j = 0; : : : ; n aj = bj+1 + tbj+2 + + tn?j bn+1 : It follows that vp(aj ) !j+1(P ) for any j in N . Thus, for any given k in N , if j k, then vp(aj ) !j+1(P ) !k+1(P );
and consequently,

!k (Q) !k+1(P ):
53

Our next corollary is a straightforward extension of the previous lemma.

Corollary 8.1. Let Q be a polynomial with rational coe cients, and assume
that t1 ; t2 ; : : : ; tk are in Z and let

P (x) = (x ? t1 )(x ? t2 )
Then !k (P ) !0 (Q).

(x ? tk )Q(x):

The main argument in our proof of Skolem's theorem is the following lemma.
sequence de ned by

Lemma 8.5. Let (dn)n2N be any sequence of integers and let (bn)n2N be the
bn =
n X n i i p di i=0

where p is an odd prime number. If bn = 0 for in nitely many indices n, then the sequence bn vanishes, i.e. bn = 0 for all n in N . Proof. For n in N , let

Rn(x) =
Then for t in N ,
n X i=0 n

n X

dipi x(x ? 1) i!(x ? i + 1) : i=0

n (t ? i + 1) = X d pi t! i i! i!(t ? i)! i=0

Rn (t) =
= and since

dipi t(t ? 1)

?t
i

X t i i p di
i=0

= 0 for i > t, it follows that

bt = Rt(t) = Rn(t) for n t:

Next, we show that for all k; n 0,

(8.3)

p?2 !k (Rn) k p ? 1 :
54

For this we write Rn(x) = n=0 c(in) xk . Clearly each c(kn) is a linear comi bination, with integer coe cients, of the numbers di pi!i , for indices i with k i n, i.e k k+1 n c(kn) = ak dk p ! + ak+1dk+1 (kp+ 1)! + + andn p ! ; k n where aj 's are integers. Consequently, i vp(c(kn) ) k inf n vp di p! ; i i and so Lemma 8.3 implies that p?1 p?2 vp(c(kn)) k inf n i p ? 1 k p ? 2 ; i which in turn shows that p?2 !k (Rn(x)) k p ? 1 : (8.4) Consider now any coe cient bt of the sequence (bn )n2N . We shall see that, for any integer k, p?2 vp(bt ) k p ? 1 ; which means that bt = 0. For this, let t1 < t2 < < tk be the rst k indices with bt1 = = btk = 0, and let n maxft; tk g. By equation (8.3), Rn(ti ) = bti = 0 for i = 1; : : : ; k. Therefore

Rn(x) = (x ? t1)(x ? t2 ) (x ? tk )Q(x) for some polynomial Q(x) with integer coe cients. By Corollary 8.1 we know that !k (Rn) !0(Q). Now vp(Rn(t)) vp(Q(t)), and by the equation (8.2), vp(Q(t)) !0(Q). So we get vp(Rn(t)) vp(Q(t)) !0(Q) !k (Rn): Finally, by equation (8.3), vp(bt ) = vp(Rn(t)), and therefore it follows from equation (8.4) that p?2 vp(bt ) k p ? 1 for all k 0. This proves the claim.
55

We shall rst prove Theorem 8.2 for regular rational series with coe cients in Z. Recall that, for a prime p, Zp = f0; 1; : : : ; p ? 1g is the ring, where the usual operations of the semiring Z are done modulo p. Zp is a ring, since it is a semiring and furthermore, every element a of Zp has an inverse element a?1 in Zp such that a a?1 = 1.

Lemma 8.6. Let S = P anxn in Z x]] be a regular rational series and let

( ; ; ) be a linear representation of S of minimal dimension k with integer coe cients. For any odd prime p not dividing det( (x)), the annihilator ann(S ) is quasiperiodic of period at most pk2 .

Proof. Let p be an odd prime not dividing det( (x)). Let n 7! n be the canonical morphism from Z to Zp, i.e. n is in f0; 1; : : : ; p ? 1g and n n (mod p). Since det( (x)) = det( (x)) 6= 0, the matrix (x) has the inverse matrix in Mk (Zp), and it can be calculated from the inverse matrix of (x) in Mk (Q ) by mapping every element to Zp. Now there are pk2 di erent elements in Mk (Zp), therefore there exist some i and j in N , i 6= j , such that i j (x) = (x) . Since (x)2 has an inverse matrix, it follows that there exists an integer N , 0 N pk , such that

(xN ) = I: This means that for the original matrix (x), there exists a matrix M with integer coe cients such that (xN ) = I + pM: Consider a xed integer j 2 f0; 1; : : : ; N ? 1g, and de ne series

bn = aj+nN ;
where n 0. Then

(8.5)

bn =
since

(xj+nN )

(xj )(I + pM )n

n X n i j i = i p (x )M ; i=0

(I + pM )n

n n X n i i X n i I n?i = = i pM : i (pM ) i=0 i=0

56

By setting di =

(xj )M i , we obtain

bn =

n X n i i p di : i=0

Now by Lemma 8.5, the sequence bn either vanishes or contains only nitely many vanishing terms. Thus the annihilator of S is a union of nite and purely periodic sets by de nition (8.5) of series bn . Therefore ann(S ) is quasiperiodic.
Proof of Theorem 8.2. Let ( ; ; ) be a linear representation of S = anxn in Z x]]. If det( (x)) 6= 0, by Lemma 8.6, ann(S ) is quasiperiodic. So we can assume that det( (x)) = 0. Let c( ) be the characteristic polynomial of (x), c( ) = (?1)k ( k + c1 k?1 + + ck ): As mentioned in the previous section, the equation an+k + c1an+k?1 + + ck an = 0; (8.6) where n 0, is a linear recurrence relation satis ed by S . Since c(0) = det( (x)), by de nition of the characteristic polynomial, necessarily ck = 0, and we may write equation (8.6) in the form an+k + c1 an+k?1 + + cm an+k?m = 0; P where 0 m < k and cm 6= 0. Let T = enxn be the rational series de ned by the (proper) linear recurrence relation en+h + c1en+h?1 + + cmen = 0; for n 0, and let the starting conditions be e0 = ak?m; e1 = ak?m+1; : : : ; eh?1 = ak?1: Now T has a linear representation ( 1 ; 1; 1) of Lemma 8.1, and det( 1 ) = cm and T is regular, and so by Lemma 8.6, ann(T ) is quasiperiodic. The annihilator of S is not necessarily quasiperiodic, but since ann(S ) = fi j 0 i < k ? m; ai = 0g (ann(T ) + (k ? m)); where ann(T ) + (k ? m) denotes the addition of (k ? m) to all elements of ann(T ), the annihilator of S is clearly a union of a nite set and of a nite number of arithmetic progressions.

57

Then the set

Corollary 8.2. Let S = P anxn be a rational series with coe cients in Q .

fn 2 N j an = 0g
is a union of a nite set and of a nite number of arithmetic progressions.

From the proof of Skolem's theorem we obtain a corollary for the rational series over rational numbers.

Proof. Let ( ; ; ) be a linear representation of S = anxn 2 Q x]], and let q be the common multiple of the denominators of coe cients in , and . Then (q ; q ; q ) is P linear representation of rational series S 0 with a 0 = q n+2 an xn , and therefore ann(S ) = ann(S 0 ). coe cients in Z. Now S

As mentioned earlier the idea of the proof in this section is due to G. Hansel. This proof, as we saw, is elementary, but not short. Skolem's original proof and its extensions by Mahler to rational series with coe cients in algebraic number elds, and by Lech to elds of characteristic 0, were much more di cult using p-adic analysis. The last extension to rational series with coe cients in a eld of characteristic 0 is referred to as Skolem{Mahler{Lech theorem. The proofs in this section were also constructive. This is very nice from our point of view, since we are looking for an algorithm to decide whether square a matrix M has in nitely many powers k such that (M k )1n = 0.

8.3 Decidability of the in nite number of zeros

We end this chapter with a theorem the proof of which is based on the constructive ideas in previous section.

Theorem 8.3. Given a square matrix M in Mn (Z), it is decidable whether

Proof. We construct a rational series S = ak xk , where ak = (M k )1n . Clearly this has an linear representation ( ; ; ), where

there exists an in nite number of natural numbers k such that (M k )1n = 0, i.e. M k has zero in the right upper corner.

= (1; 0; : : : ; 0) 2 Z1 n;

(x) = M and 58

= (0; : : : ; 0; 1)T 2 Zn 1:

First we decide whether (x) is invertible or not, i.e. whether the linear representation ( ; ; ) is regular. If not, we consider the regular rational series T de ned in the proof of Skolem's theorem. Now we have a regular series to consider, and we can use the proof of Lemma 8.5. By this proof we can compute the period N , and we then divide the rational series considered into series (bn)n 0. We obtain N series, and from the proof we also get linear representations satis ed by these series. So we can compute linear recurrence relations satis ed by these series from the characteristic polynomials of their linear representations. Finally we know that these series (bn)n 0 either vanish or have only nitely many nonzero elements. If some of these series vanish then there is in nitely many n such that (M k )1n = 0, otherwise none.

59

9 Summary
We have considered many simply formulated problems in matrix theory. Some of them are proved to be decidable and some undecidable, but many problems are also left open. For instance, the freeness and the mortality problem for 2 2 matrices and Skolem's problem for n n matrices, where n 3, were such. We have also seen that the Post Correspondence Problem is very useful in the proofs of the undecidability results in the matrix theory. Actually all undecidability results were based on the undecidability of PCP. They also used the same method, coding of two independent words to matrices, a method that was rst used by Paterson. We also notice that the proofs for the decidable cases are complicated, and need some backround from other branches of mathematics. We used combinatorics, graph theory and theory of rational series in these proofs. We shall now present the results proved in this work in the form of a table. The entries are vectors of two parameters (n; k), where n is the dimension of matrices and k the number of matrices. We restrict to the cases, where matrices are over integers. Numbers n and k in the table present arbitrary values of these parameters and Skolem's problem is included in the problem of the zero in the right upper corner.
Problem

Mortality (2; 2), (n; 1) ( 3; 15), ( 45; 2) Freeness (n; 1) ( 3; 18) Finiteness (n; k) Zero in the right upper corner ( 2; 1) ( 3; 7), ( 24; 2) Common element ( 3; 7) Table: Summary of decidable and undecidable matrix problems.

Decidable

Undecidable

60

References
BeP] J. Berstel and D. Perrin, Theory of Codes. Academic Press, 1986. BeR] J. Berstel and C. Reutenauer, Rational Series and Their Languages, Springer-Verlag, 1984. Cas] J. Cassaigne, personal communication. CHK] J. Cassaigne, T. Harju and J. Karhumaki, On the decidability of the freeness of matrix semigroups, TUCS Technical Report No 56. CKa] J. Cassaigne and J. Karhumaki, Examples of undecidable problems for 2-generator matrix semigroups, TUCS Technical Report No 57. Gan] F. R. Gantmacher, The Theory of Matrices, Vol. I. Chelsea Publishing Company, New York, 1959. GRS] R. Graham, R Rothschild and J. Spencer, Ramsey Theory, John Wiley & Sons, 1990. HaK] T. Harju and J. Karhumaki, Morphisms, In G. Rozenberg and A. Salomaa, editors, Handbook of Formal Languages, Springer Verlag, 1997. Han] G. Hansel, Une demonstration simple du theoreme de Skolem-MahlerLech, Theoret. Comput. Sci. 43 (1986), 1-10. Ja1] G. Jacob, La nitude des representations lineaires des semi-groupes est decidable, J. Algebra 52 (1978), 437-459. Ja2] G. Jacob, Un algorithme calculant le cardinal, ni ou in ni, des demi groups de matrices, Theoret. Comput. Sci. 5 (1977). KBS] D.A. Klarner, J.-C. Birget and W. Satter eld, On the undecidability of the freeness of the integer matrix semigroups, Int. J. Algebra Comp. 1 (1991), 223-226. Kro] M. Krom, An unsolvable problem with product of matrices, Math. System Theory 14 (1981), 335-337. MaS] A. Mandel and I. Simon, On nite semigroups of matrices, Theoret. Comput. Sci. 5 (1977), 101-111. Man] Z. Manna, Mathematical Theory of Computations, McGraw-Hill, 1974. 61

MSe] Y. Matiyasevich and G. Senizerques, Decision problems for semi Thue systems with a few rules. Manuscript, 1996. Pat] M.S. Paterson, Unsolvability in 3 3 matrices, Studies in Appl. Math. 49 (1970), 105-107. Pos] E. Post, A variant of a recursively unsolvable problem, Bulletin of Amer. Math. Soc. 52 (1946), 264-268. Ram] F.P. Ramsey, On a problem of formal logic, Proc. London Math. Soc. 2nd Ser. 30 (1930), 264-286. Sch] P. Schultz, Mortality of 2 2 matrices, Amer. Math. Monthly 84 (1977), 463-464. Scb] M.P. Schutzenberger, On the de nition of a family of automata, Information and Control 4 (1961), 245-270. WeS] A. Weber and H. Seidl, On nitely generated monoids of matrices with entries in N , RAIRO Inform. Theor., 25 (1991), 19-38.

62

Turku Centre for Computer Science Lemminkaisenkatu 14 FIN-20520 Turku Finland

http://www.tucs.abo.

University of Turku Department of Mathematical Sciences

Abo Akademi University Department of Computer Science Institute for Advanced Management Systems Research

Turku School of Economics and Business Administration Institute of Information Systems Science